WO2023055153A1 - Method, device and recording medium for image encoding/decoding - Google Patents

Abstract

A method, a device and a recording medium for image encoding/decoding are disclosed. Provided in embodiments are a method, a device and a recording medium for efficiently lightening an in-loop filter and a predictive deep neural network in order to enhance encoding efficiency of an image and shorten the time required for encoding. Provided in embodiments is learning in a lightweight in-loop filter and a lightweight predictive deep neural network through knowledge distillation. According to embodiments, provided is a method for learning in an in-loop filter and a predictive deep neural network in order to enhance the efficiency and speed of image encoding and shorten the time required for encoding.

Description

Method, apparatus and recording medium for video encoding/decoding

The present invention relates to a method, apparatus and recording medium for video encoding/decoding.

The present invention is the benefit of the filing date of Korean Patent Application No. 10-2021-0129157 filed on September 29, 2021, the benefit of the filing date of Korean Patent Application No. 10-2022-0107154 filed on August 25, 2022, and the Claims the benefit of the filing date of Korean Patent Application No. 10-2022-0124695 filed on the 29th of the month, all of which are incorporated herein.

Through the continuous development of the information communication industry, broadcasting services having high definition (HD) resolution have been spread worldwide. Through this proliferation, many users have become accustomed to high-resolution and high-definition images and/or videos.

In order to satisfy users' demand for high image quality, many organizations are accelerating the development of next-generation video devices. Users' interest in Ultra High Definition (UHD) TV, which has four times higher resolution than FHD TV, as well as High Definition TV (HDTV) and Full HD (FHD) TV. has increased, and according to this increase in interest, an image encoding/decoding technology for an image having a higher resolution and quality is required.

As video compression technology, there are various technologies such as inter prediction technology, intra prediction technology, transformation and quantization technology, and entropy encoding technology.

The inter-prediction technique is a technique of predicting a value of a pixel included in a current picture using pictures before and/or after the current picture. Intra-prediction technology is a technology of predicting a value of a pixel included in a current picture by using information about pixels in the current picture. The transformation and quantization technique is a technique for compressing the energy of the residual image. The entropy coding technique assigns short codes to values with a high frequency of occurrence and assigns long codes to values with a low frequency of occurrence.

Using this video compression technology, data for video can be effectively compressed, transmitted, and stored.

An embodiment may provide a method, apparatus, and recording medium for providing learning in an in-loop filter and predictive deep neural network.

In one aspect, selecting a deep neural network; generating coded deep neural network information by encoding deep neural network information used to construct the deep neural network; and generating a bitstream including the encoded deep neural network information.

The image encoding method may further include performing learning in the deep neural network.

The step of performing learning in the deep neural network may include: performing learning in a teacher network; conducting first learning in the student network; and performing second learning in the student network.

Learning in the student network may be performed using knowledge distillation using the teacher network.

Learning in an in-loop filter based on the deep neural network may be performed.

An input of the in-loop filter may be an image in which compression degradation occurs.

An output of the in-loop filter may be an image from which compression degradation has been removed.

Learning in the in-loop filter may be performed to minimize an error between the output of the in-loop filter and the original image.

The deep neural network may be a predictive deep neural network that generates a predicted image similar to an original image through prediction.

In another aspect, receiving a bitstream including coded deep neural network information; generating deep neural network information used to construct a deep neural network by decoding the encoded deep neural network information; and selecting the deep neural network.

The image decoding method may further include performing learning in the deep neural network.

The step of performing learning in the deep neural network may include: performing learning in a teacher network; conducting first learning in the student network; and performing second learning in the student network.

Learning in the student network may be performed using knowledge distillation using the teacher network.

Learning in an in-loop filter based on the deep neural network may be performed.

An input of the in-loop filter may be an image in which compression degradation occurs.

An output of the in-loop filter may be an image from which compression degradation has been removed.

Learning in the in-loop filter may be performed to minimize an error between the output of the in-loop filter and the original image.

The deep neural network may be a predictive deep neural network that generates a predicted image similar to an original image through prediction.

In another aspect, in a computer-readable recording medium storing a bitstream for image decoding, the bitstream includes encoded deep neural network information, and decoding the encoded deep neural network information is performed to A computer readable recording medium in which deep neural network information used to construct a deep neural network is generated and the deep neural network is selected is provided.

Learning in the deep neural network may be performed.

Learning in the teacher network can be performed.

A first learning in the student network may be performed.

Second learning in the student network may be performed.

Learning in the student network may be performed using knowledge distillation using the teacher network.

Learning in an in-loop filter based on the deep neural network may be performed.

An input of the in-loop filter may be an image in which compression degradation occurs.

An output of the in-loop filter may be an image from which compression degradation has been removed.

Learning in the in-loop filter may be performed to minimize an error between the output of the in-loop filter and the original image.

A method, apparatus, and recording medium for providing learning in an in-loop filter and predictive deep neural network are provided.

1 is a block diagram showing a configuration according to an embodiment of an encoding device to which the present invention is applied.

2 is a block diagram showing a configuration according to an embodiment of a decoding device to which the present invention is applied.

3 is a diagram schematically illustrating a division structure of an image when encoding and decoding an image.

4 is a diagram illustrating a form of a prediction unit that a coding unit may include.

5 is a diagram illustrating a form of a transform unit that may be included in a coding unit.

6 shows division of a block according to an example.

7 is a diagram for explaining an embodiment of an intra prediction process.

8 is a diagram for explaining reference samples used in an intra prediction process.

9 is a diagram for explaining an embodiment of an inter prediction process.

10 shows spatial candidates according to an example.

11 illustrates an order of adding motion information of spatial candidates to a merge list according to an example.

12 illustrates a process of transformation and quantization according to an example.

13 illustrates diagonal scanning according to an example.

14 illustrates horizontal scanning according to an example.

15 illustrates vertical scanning according to an example.

16 is a structural diagram of an encoding device according to an embodiment.

17 is a structural diagram of a decryption device according to an embodiment.

18 is a flowchart of an encoding method according to an embodiment.

19 is a flowchart of a decoding method according to an embodiment.

20 is a flowchart of learning in a deep neural network according to an example.

21 illustrates learning in a teacher network according to an example.

22 shows learnings in an intra-prediction deep neural network and an inter-prediction deep neural network according to an embodiment.

23 illustrates a first learning in a student network according to one embodiment.

24 illustrates a first learning in a student network according to one embodiment.

25 illustrates second learning in a student network according to one embodiment.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS For detailed descriptions of exemplary embodiments described below, reference is made to the accompanying drawings, which illustrate specific embodiments by way of example. These embodiments are described in sufficient detail to enable those skilled in the art to practice the embodiments. It should be understood that the various embodiments are different, but need not be mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in one embodiment in another embodiment without departing from the spirit and scope of the invention. Additionally, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the embodiment. Accordingly, the detailed description set forth below is not to be taken in a limiting sense, and the scope of the exemplary embodiments, if properly described, is limited only by the appended claims, along with all equivalents as claimed by those claims.

Like reference numbers in the drawings indicate the same or similar function throughout the various aspects. The shapes and sizes of elements in the drawings may be exaggerated for clarity.

In the present invention, terms such as first and second may be used to describe various components, but the components should not be limited by the terms. These terms are only used for the purpose of distinguishing one component from another. For example, a first element may be termed a second element, and similarly, a second element may be termed a first element, without departing from the scope of the present invention. The term "and/or" may include any combination of a plurality of related listed items or any of a plurality of related listed items.

When a component is referred to as being “connected” or “connected” to another component, the two components may be directly connected or connected to each other, but It will be understood that other components may exist in the middle of the two components. On the other hand, when a component is referred to as "directly connected" or "directly connected" to another component, it should be understood that no other component exists in the middle of the above two components. something to do.

Elements appearing in the embodiments are shown independently to represent different characteristic functions, and do not mean that each element is composed of separate hardware or a single software unit. That is, each component is listed and included as each component for convenience of description, and at least two of each component are combined to form one component, or one component is divided into a plurality of components to perform functions. It can be performed, and integrated embodiments and separate embodiments of each of these components are also included in the scope of the present invention as long as they do not depart from the essence of the present invention.

Terms used in the embodiments are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In embodiments, the terms "comprise" or "have" are intended to indicate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other It should be understood that the presence or addition of features, numbers, steps, operations, components, parts, or combinations thereof is not precluded. That is, the description of "including" a specific configuration in the embodiments does not exclude configurations other than the corresponding configuration, and means that additional configurations may also be included in the practice of the present invention or the scope of the technical idea of the present invention. .

In embodiments, the term “at least one” may mean one or more numbers such as 1, 2, 3, and 4. In embodiments, the term “a plurality of” may mean one of two or more numbers, such as 2, 3, and 4.

Some of the components of the embodiments are not essential components that perform essential functions in the present invention, but may be optional components for improving performance. Embodiments may be implemented by including only essential components in implementing the essence of the embodiments, excluding components used for performance improvement. A structure including only essential components excluding optional components used for performance improvement is also included in the scope of the embodiments.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily practice the embodiments. In describing the embodiments, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present specification, the detailed description will be omitted. In addition, the same reference numerals are used for the same components in the drawings, and duplicate descriptions of the same components are omitted.

Hereinafter, an image may mean one picture constituting a video, and may also indicate the video itself. For example, "encoding and/or decoding an image" may mean "encoding and/or decoding a video", and may mean "encoding and/or decoding one of images constituting a video". may be

Hereinafter, the terms “video” and “motion picture(s)” may be used interchangeably and may be used interchangeably.

Hereinafter, the target image may be an encoding target image that is an encoding target and/or a decoding target image that is a decoding target. Also, the target image may be an input image input to an encoding device or an input image input to a decoding device. Also, the target image may be a current image that is a target of current encoding and/or decoding. For example, the terms “target image” and “current image” may be used interchangeably and may be used interchangeably.

Hereinafter, the terms “image”, “picture”, “frame” and “screen” may be used interchangeably and interchangeably.

Hereinafter, the target block may be an encoding target block that is an encoding target and/or a decoding target block that is a decoding target. Also, the target block may be a current block that is a target of current encoding and/or decoding. For example, the terms “target block” and “current block” may be used interchangeably and may be used interchangeably. The current block may refer to a coding target block that is an encoding target during encoding and/or a decoding target block that is a decoding target during decoding. Also, the current block may be at least one of a coding block, a prediction block, a residual block, and a transform block.

Hereinafter, the terms “block” and “unit” may be used interchangeably and may be used interchangeably. Or "block" may represent a specific unit.

In the following, the terms "region" and "segment" may be used interchangeably.

In embodiments, each of the specified information, data, flag, index and element, attribute, etc. may have a value. The value "0" of information, data, flags, indexes, elements, and attributes may represent false, logical false, or a first predefined value. That is to say, the value "0", false, logic false and the first predefined value may be used interchangeably. The value "1" of information, data, flags, indexes, elements, and attributes may represent true, logical true, or a second predefined value. In other words, the value "1", true, logically true, and the second predefined value may be used interchangeably.

When a variable such as i or j is used to indicate a row, column, or index, the value of i may be an integer greater than or equal to 0, or may be an integer greater than or equal to 1. That is to say, row, column, index, etc. may be counted from 0, in embodiments, may be counted from 1.

In embodiments, the term “one or more” or the term “at least one” may mean the term “plurality”. “One or more” or “at least one” may be used interchangeably with “plural”.

Below, terms used in the embodiments are explained.

Encoder: An encoder may mean a device that performs encoding. In other words, an encoder may mean an encoding device.

Decoder: A decoder may refer to a device that performs decoding. In other words, a decryptor may mean a decryption device.

Unit: A unit may represent a unit of encoding and/or decoding of an image. The terms "unit" and "block" may be used interchangeably and may be used interchangeably.

- A unit may be an MxN array of samples. M and N may each be a positive integer. A unit may refer to an array of samples in a two-dimensional form.

- In encoding and decoding of an image, a unit may be a region created by dividing one image. In other words, a unit may be a specified area within one image. One image may be divided into a plurality of units. Alternatively, the unit may refer to the divided parts when one image is divided into subdivided parts and encoding or decoding of the divided parts is performed.

- In encoding and decoding of an image, a predefined process for a unit may be performed according to a unit type.

- Depending on the function, the type of unit is a macro unit, a coding unit (CU), a prediction unit (PU), a residual unit, and a transform unit (TU), etc. can be classified as Or, depending on the function, the unit is a block, macroblock, coding tree unit, coding tree block, coding unit, coding block, or prediction unit. It may mean a prediction unit, a prediction block, a residual unit, a residual block, a transform unit, a transform block, and the like. For example, the target unit may be at least one of a CU, a PU, a residual unit, and a TU to be encoded and/or decoded.

- A unit may refer to information including a luma component block, a chroma component block corresponding to the luma component block, and a syntax element for each block, in order to be referred to as a block.

- Units may vary in size and shape. Also, units can have various sizes and shapes. In particular, the shape of the unit may include not only a square but also a two-dimensional geometric figure such as a rectangle, a trapezoid, a triangle, and a pentagon.

- In addition, the unit information may include at least one or more of a unit type, a size of a unit, a depth of a unit, a coding order of a unit, a decoding order of a unit, and the like. For example, the unit type may indicate one of CU, PU, residual unit, and TU.

- One unit can be further divided into sub-units having a smaller size than the unit.

Depth: Depth may mean the degree of division of a unit. Also, the depth of a unit may indicate a level at which the unit exists when the unit(s) are expressed as a tree structure.

- Unit division information may include depth about the depth of the unit. Depth may indicate the number and/or extent to which a unit is divided.

- In the tree structure, the root node has the shallowest depth and the leaf node has the deepest depth. The root node may be the highest node. A leaf node may be the lowest node.

- One unit may be hierarchically divided into a plurality of sub-units while having depth information based on a tree structure. In other words, a unit and a sub-unit generated by division of the unit may correspond to a node and a child node of the node, respectively. Each divided sub-unit may have a depth. Since depth represents the number and/or degree of division of a unit, division information of a sub-unit may include information about the size of the sub-unit.

- In the tree structure, the highest node may correspond to the first non-split unit. The highest node may be referred to as a root node. Also, the highest node may have the smallest depth value. In this case, the highest node may have a depth of level 0.

- A node with a depth of level 1 may represent a unit created as the original unit is split once. A node with a depth of level 2 may represent a unit created as the original unit is split twice.

- A node with a depth of level n may represent a unit created as the initial unit is split n times.

- A leaf node may be the lowest node and may be a node that cannot be further divided. The depth of a leaf node may be the maximum level. For example, the predefined value of the maximum level may be 3.

-QT depth may indicate the depth for quad division. BT depth may indicate a depth for binary partitioning. The TT depth may indicate a depth for ternary division.

Sample: A sample may be a base unit constituting a block. A sample may be represented as values from 0 to 2 ^Bd -1 according to a bit depth (Bd).

- A sample can be a pixel or a pixel value.

- In the following, the terms "pixel", "pixel" and "sample" may be used in the same meaning and may be used interchangeably.

Coding Tree Unit (CTU): A CTU may consist of one luma component (Y) coding tree block and two chroma component (Cb, Cr) coding tree blocks related to the luma component coding tree block. there is. In addition, the CTU may mean including syntax elements for the above blocks and each block of the above blocks.

- Each coding tree unit is a quad tree (QT), binary tree (BT), and ternary tree (TT) to construct sub units such as a coding unit, a prediction unit, and a transform unit. It can be segmented using one or more segmentation schemes. A quad tree may mean a quarternary tree. In addition, each coding tree unit may be split using a MultiType Tree (MTT) using one or more splitting schemes.

- CTU may be used as a term to refer to a pixel block, which is a processing unit in the process of decoding and encoding an image, as in segmentation of an input image.

Coding Tree Block (CTB): A coding tree block may be used as a term to refer to any one of a Y coding tree block, a Cb coding tree block, and a Cr coding tree block.

Neighbor Block: A neighboring block may mean a block adjacent to a target block. A neighboring block may mean a reconstructed neighboring block.

- In the following, the terms "neighboring block" and "adjacent block" may be used in the same meaning and may be used interchangeably.

- A neighboring block may mean a reconstructed neighboring block.

Spatial neighbor block: A spatial neighbor block may be a block that is spatially adjacent to the target block. Neighboring blocks may include spatial neighboring blocks.

- The target block and spatial neighboring blocks may be included in the target picture.

- A spatial neighboring block may mean a block whose boundary meets the target block or a block located within a predetermined distance from the target block.

- A spatial neighboring block may mean a block adjacent to a vertex of a target block. Here, the block adjacent to the vertex of the target block may be a block vertically adjacent to a neighboring block horizontally adjacent to the target block or a block horizontally adjacent to a neighboring block vertically adjacent to the target block.

Temporal neighbor block: A temporal neighbor block may be a block that is temporally adjacent to the target block. Neighboring blocks may include temporal neighboring blocks.

- A temporal neighboring block may include a co-located block (col block).

- A collocated block may be a block in an already reconstructed co-located picture (col picture). A position of a collocated block in a collocated picture may correspond to a position of a target block in a target picture. Alternatively, the position of the collocated block in the collocated picture may be the same as the position of the target block in the target picture. A collocated picture may be a picture included in a reference picture list.

- A temporal neighboring block may be a block temporally adjacent to a spatial neighboring block of a target block.

Prediction mode: The prediction mode may be information indicating a mode used for intra prediction or a mode used for inter prediction.

Prediction unit: A prediction unit may mean a base unit for prediction such as inter prediction, intra prediction, inter compensation, intra compensation, and motion compensation.

- One prediction unit may be divided into a plurality of smaller-sized partitions or sub-prediction units. A plurality of partitions may also be a basis unit in performing prediction or compensation. A partition generated by division of a prediction unit may also be a prediction unit.

Prediction unit partition (prediction unit partition): A prediction unit partition may mean a form in which a prediction unit is divided.

Reconstructed neighboring unit: A reconstructed neighboring unit may be a unit that has already been decoded and reconstructed in a neighbor of a target unit.

- The reconstructed neighbor unit may be a spatial neighbor unit or a temporal neighbor unit to the target unit.

- The reconstructed spatial neighboring unit may be a unit in the target picture and already reconstructed through encoding and/or decoding.

- The reconstructed temporal neighbor unit may be a unit in the reference picture and a unit that has already been reconstructed through encoding and/or decoding. A position of the reconstructed temporal neighboring unit within the reference image may be the same as a position within the target unit's target picture or may correspond to a position within the target unit's target picture. Alternatively, the reconstructed temporal neighboring unit may be a neighboring block of a corresponding block in the reference picture. Here, the position of the corresponding block in the reference image may correspond to the position of the target block in the target image. Here, the correspondence of the positions of the blocks may mean that the positions of the blocks are the same, and may mean that one block is included in another block, and one block occupies a specified position of the other block. can mean doing

Sub-picture: A picture can be divided into one or more sub-pictures. A sub-picture may consist of one or more tile rows and one or more tile columns.

- A sub-picture may be an area having a square shape or a rectangular (ie, non-square) shape within a picture. In addition, a sub-picture may include one or more CTUs. .

- A sub-picture may be a rectangular area of one or more slices within one picture.

- One sub-picture may include one or more tiles, one or more bricks, and/or one or more slices.

Tile: A tile can be a square or rectangular (ie, non-square) area within a picture.

- A tile may contain one or more CTUs.

- A tile can be divided into one or more bricks.

Brick: A brick may mean one or more CTU rows within a tile.

- A tile can be divided into one or more bricks. Each brick may contain one or more CTU rows.

- A tile that is not divided into two or more can also mean a brick.

Slice: A slice can include one or more tiles within a picture. Or, a slice may include one or more bricks within a tile.

- A sub-picture may contain one or more slices that collectively cover a rectangular area within the picture. Accordingly, each sub-picture boundary may always be a slice boundary. Also, each vertical sub-picture boundary may always be a vertical tile boundary.

Parameter set: A parameter set may correspond to header information among structures in a bitstream.

- Parameter sets include a Video Parameter Set (VPS), a Sequence Parameter Set (SPS), a Picture Parameter Set (PPS), an Adaptation Parameter Set (APS), and a decoding parameter It may include at least one of a set (Decoding Parameter Set; DPS), and the like.

Information signaled through a parameter set may be applied to pictures referring to the parameter set. For example, information in the VPS may be applied to pictures referring to the VPS. Information in the SPS may be applied to pictures referring to the SPS. Information in the PPS may be applied to pictures referring to the PPS.

A parameter set may refer to an upper parameter set. For example, PPS may refer to SPS. SPS may refer to VPS.

- In addition, the parameter set may include tile group, slice header information, and tile header information. A tile group may refer to a group including a plurality of tiles. Also, the meaning of a tile group may be the same as that of a slice.

Rate-distortion optimization: The encoding device uses a combination of the size of the coding unit, the prediction mode, the size of the prediction unit, motion information, and the size of the conversion unit to provide high encoding efficiency. Distortion optimization can be used.

- The rate-distortion optimization method may calculate a rate-distortion cost of each combination in order to select an optimal combination among the above combinations. The rate-distortion cost can be calculated using the formula “D+λ*R”. In general, a combination that minimizes the rate-distortion cost according to the equation “D+λ*R” can be selected as an optimal combination in the rate-distortion optimization method.

- D can represent distortion. D may be the mean square error of difference values between the original transform coefficients and the reconstructed transform coefficients within the transform unit.

- R can represent rate. R may represent a bit rate using related context information.

- λ may represent a Lagrangian multiplier. R may include not only coding parameter information such as a prediction mode, motion information, and a coded block flag, but also bits generated by encoding transform coefficients.

- The encoding device may perform processes such as inter prediction, intra prediction, transformation, quantization, entropy encoding, inverse quantization, and/or inverse transformation to calculate accurate D and R. These processes can greatly increase complexity in an encoding device.

Bitstream: A bitstream may mean a string of bits including coded image information.

Parsing: Parsing may mean determining a value of a syntax element by entropy decoding a bitstream. Alternatively, parsing may mean entropy decoding itself.

Symbol: may mean at least one of a syntax element, a coding parameter, and a transform coefficient of a coding target unit and/or a decoding target unit. Also, a symbol may mean an object of entropy encoding or a result of entropy decoding.

Reference picture: A reference picture may refer to an image that a unit refers to for inter prediction or motion compensation. Alternatively, the reference picture may be an image including a reference unit referred to by a target unit for inter prediction or motion compensation.

Hereinafter, the terms “reference picture” and “reference image” may be used interchangeably and may be used interchangeably.

Reference picture list: The reference picture list may be a list including one or more reference pictures used for inter prediction or motion compensation.

- The types of reference picture lists are List Combined (LC), List 0 (L0), List 1 (L1), List 2 (List 2; L2), and List 3 (List 3; L3). ), etc. may be present.

- One or more reference picture lists may be used for inter prediction.

Inter prediction indicator: The inter prediction indicator may indicate the direction of inter prediction for a target unit. Inter prediction can be one of uni-prediction and bi-prediction, etc. Alternatively, the inter prediction indicator may indicate the number of reference pictures used when generating a prediction unit of a target unit. Alternatively, the inter prediction indicator may indicate the number of prediction blocks used for inter prediction or motion compensation of the target unit.

Prediction list utilization flag: The prediction list utilization flag may indicate whether a prediction unit is generated using at least one reference picture in a specific reference picture list.

- An inter prediction indicator can be derived using the prediction list utilization flag. Conversely, the prediction list utilization flag can be derived using the inter prediction indicator. For example, when the prediction list utilization flag indicates a first value of 0, it may indicate that a prediction block is not generated using a reference picture in the reference picture list for the target unit. When the prediction list utilization flag indicates the second value of 1, it may indicate that a prediction unit is generated using the reference picture list for the target unit.

Reference picture index: The reference picture index may be an index indicating a specific reference picture in a reference picture list.

Picture Order Count (POC): The POC of a picture may indicate a display order of pictures.

Motion Vector (MV): A motion vector may be a two-dimensional vector used in inter prediction or motion compensation. A motion vector may mean an offset between a target image and a reference image.

- For example, MV can be expressed as (mv _x , mv _y ). mv _x may represent a horizontal component, and mv _y may represent a vertical component.

Search range: The search range may be a two-dimensional area where MVs are searched during inter prediction. For example, the size of the search area may be MxN. M and N may each be a positive integer.

Motion vector candidate: A motion vector candidate may mean a block as a prediction candidate or a motion vector of a block as a prediction candidate when predicting a motion vector.

- A motion vector candidate may be included in a motion vector candidate list.

Motion vector candidate list: The motion vector candidate list may refer to a list constructed using one or more motion vector candidates.

Motion vector candidate index: The motion vector candidate index may mean an indicator indicating a motion vector candidate in the motion vector candidate list. Alternatively, the motion vector candidate index may be an index of a motion vector predictor.

Motion information: Motion information includes not only motion vectors, reference picture indices and inter prediction indicators, but also reference picture list information, reference pictures, motion vector candidates, motion vector candidate indices, merge candidates and merge indices, etc. It may mean information including at least one of

Merge candidate list: A merge candidate list may refer to a list constructed using one or more merge candidates.

Merge candidate: A merge candidate is a spatial merge candidate, a temporal merge candidate, a combined merge candidate, a combined bi-prediction merge candidate, a candidate based on history, a candidate based on an average of two candidates, and zero It may mean a merge candidate and the like. A merge candidate may include an inter prediction indicator, and may include motion information such as a reference picture index for each list, a motion vector, a prediction list utilization flag, and an inter prediction indicator.

Merge index: A merge index may be an indicator pointing to a merge candidate in a merge candidate list.

- The merge index may indicate a reconstructed unit that derives a merge candidate from among reconstructed units spatially adjacent to the target unit and reconstructed units temporally adjacent to the target unit.

- The merge index may indicate at least one piece of motion information of a merge candidate.

Transform unit: A transform unit may be a basic unit in residual signal encoding and/or residual signal decoding, such as transform, inverse transform, quantization, inverse quantization, transform coefficient encoding, and transform coefficient decoding. One transform unit may be divided into a plurality of sub-transform units having smaller sizes. Here, the transformation may include one or more of a first-order transformation and a second-order transformation, and the inverse transformation may include one or more of a first-order inverse transformation and a second-order inverse transformation.

Scaling: Scaling may refer to a process of multiplying a transform coefficient level by a factor.

- As a result of the scaling to the transform coefficient level, a transform coefficient can be generated. Scaling may be referred to as dequantization.

Quantization Parameter (QP): A quantization parameter may mean a value used when generating a transform coefficient level for a transform coefficient in quantization. Alternatively, the quantization parameter may refer to a value used when generating a transform coefficient by scaling a transform coefficient level in inverse quantization. Alternatively, the quantization parameter may be a value mapped to a quantization step size.

Delta quantization parameter: The delta quantization parameter may refer to a difference value between a predicted quantization parameter and a quantization parameter of a target unit.

Scan: A scan may refer to a method of arranging the order of coefficients within a unit, block or matrix. For example, arranging a 2D array into a 1D array may be referred to as a scan. Alternatively, arranging a one-dimensional array into a two-dimensional array may also be referred to as scan or inverse scan.

Transform coefficient: A transform coefficient may be a coefficient value generated by performing transformation in an encoding device. Alternatively, the transform coefficient may be a coefficient value generated by performing at least one of entropy decoding and inverse quantization in the decoding apparatus.

- A quantized level generated by applying quantization to a transform coefficient or a residual signal or a quantized transform coefficient level may also be included in the meaning of a transform coefficient.

Quantized level: A quantized level may refer to a value generated by performing quantization on a transform coefficient or a residual signal in an encoding device. Alternatively, the quantized level may mean a value to be subjected to inverse quantization when the decoding apparatus performs inverse quantization.

- A quantized transform coefficient level, which is a result of transform and quantization, may also be included in the meaning of the quantized level.

Non-zero transform coefficient: A non-zero transform coefficient may mean a transform coefficient having a non-zero value or a transform coefficient level having a non-zero value. Alternatively, the non-zero transform coefficient may refer to a transform coefficient whose value is not 0 or a transform coefficient level whose value is not 0.

Quantization matrix: A quantization matrix may mean a matrix used in a quantization process or an inverse quantization process to improve subjective or objective picture quality of an image. A quantization matrix may also be referred to as a scaling list.

Quantization matrix coefficient: A quantization matrix coefficient may mean each element in a quantization matrix. Quantization matrix coefficients may also be referred to as matrix coefficients.

Default matrix: The default matrix may be a quantization matrix predefined in an encoding device and a decoding device.

Non-default matrix: The non-default matrix may be a quantization matrix that is not predefined in the encoding device and the decoding device. The non-default matrix may refer to a quantization matrix signaled from an encoding device to a decoding device by a user.

Most Probable Mode (MPM): MPM may indicate an intra prediction mode that is highly likely to be used for intra prediction of a target block.

- The encoding device and the decoding device may determine one or more MPMs based on a coding parameter related to the target block and an attribute of an object related to the target block.

- The encoding device and the decoding device may determine one or more MPMs based on the intra prediction mode of the reference block. Reference blocks may be plural. The plurality of reference blocks may include a spatial neighboring block adjacent to the left side of the target block and a spatial neighboring block adjacent to the top of the target block. In other words, one or more different MPMs may be determined depending on which intra prediction modes are used for reference blocks.

- One or more MPMs may be determined in the same way in the encoding device and the decoding device. In other words, the encoding device and the decoding device may share an MPM list including one or more identical MPMs.

MPM List: An MPM list can be a list containing one or more MPMs. The number of one or more MPMs in the MPM list may be predefined.

MPM indicator: The MPM indicator may indicate an MPM used for intra prediction of a target block among one or more MPMs in the MPM list. For example, the MPM indicator may be an index to an MPM list.

- Since the MPM list is determined in the same way in the encoding device and the decoding device, the MPM list itself may not need to be transmitted from the encoding device to the decoding device.

- The MPM indicator may be signaled from the encoding device to the decoding device. As the MPM indicator is signaled, the decoding apparatus may determine an MPM to be used for intra prediction of the target block among MPMs in the MPM list.

MPM use indicator: The MPM use indicator may indicate whether an MPM use mode is to be used for prediction of a target block. The MPM use mode may be a mode for determining an MPM to be used for intra prediction of a target block by using an MPM list.

- The MPM use indicator may be signaled from the encoding device to the decoding device.

Signaling: Signaling may indicate that information is transmitted from an encoding device to a decoding device. Alternatively, signaling may mean that an encoding device includes information in a bitstream or a recording medium. Information signaled by the encoding device may be used by the decoding device.

- The encoding device may generate encoded information by performing encoding on signaled information. Encoded information may be transmitted from an encoding device to a decoding device. The decoding apparatus may obtain information by decoding the transmitted encoded information. Here, encoding may be entropy encoding, and decoding may be entropy decoding.

Optional Signaling: Information can be selectively signaled. Selective signaling of information may mean that an encoding device selectively includes information in a bitstream or a recording medium (according to specific conditions). Selective signaling of information may mean that a decoding apparatus selectively extracts information from a bitstream (according to a specific condition).

Omission of signaling: Signaling of information may be omitted. Omission of signaling of information about information may mean that an encoding device does not include information in a bitstream or a recording medium (according to a specific condition). Omission of signaling for information may mean that the decoding apparatus does not extract information from the bitstream (according to a specific condition).

Statistical value: Variables, coding parameters and constants, etc. can have values that can be computed. A statistical value may be a value generated by an operation on the values of these specified objects. For example, the statistical value is an average value, a weighted average value, a weighted sum, a minimum value, a maximum value, and a mode for values such as a specified variable, a specified coding parameter, and a specified constant. It can be one or more of a value, a median value, and an interpolated value.

1 is a block diagram showing a configuration according to an embodiment of an encoding device to which the present invention is applied.

The encoding device 100 may be an encoder, a video encoding device, or an image encoding device. A video may include one or more images. The encoding apparatus 100 may sequentially encode one or more images of a video.

Referring to FIG. 1, an encoding apparatus 100 includes an inter prediction unit 110, an intra prediction unit 120, a switch 115, a subtractor 125, a transform unit 130, a quantization unit 140, and entropy encoding. It may include a unit 150, an inverse quantization unit 160, an inverse transform unit 170, an adder 175, a filter unit 180, and a reference picture buffer 190.

The encoding apparatus 100 may perform encoding on a target image using an intra mode and/or an inter mode. In other words, the prediction mode for the target block may be one of an intra mode and an inter mode.

Hereinafter, the terms “intra mode”, “intra prediction mode”, “in-picture mode” and “in-picture prediction mode” may be used interchangeably and interchangeably.

Hereinafter, the terms “inter mode”, “inter prediction mode”, “inter-screen mode” and “inter-prediction mode” may be used in the same meaning and may be used interchangeably.

Hereinafter, the term “video” may refer to only a part of an image or may refer to a block. Also, processing of “image” may indicate sequential processing of a plurality of blocks.

Also, the encoding device 100 may generate a bitstream including encoded information through encoding of a target image, and output and store the generated bitstream. The generated bitstream may be stored in a computer readable recording medium and may be streamed through a wired and/or wireless transmission medium.

As the prediction mode, if the intra mode is used, the switch 115 can be switched to intra. As the prediction mode, when the inter mode is used, the switch 115 can be switched to inter.

The encoding apparatus 100 may generate a prediction block for the target block. Also, after the prediction block is generated, the encoding apparatus 100 may encode a residual block for the target block by using the target block and the residual of the prediction block.

When the prediction mode is the intra mode, the intra predictor 120 may use a pixel of a block already encoded and/or decoded, which is adjacent to the target block, as a reference sample. The intra predictor 120 may perform spatial prediction on the target block using the reference sample, and generate prediction samples for the target block through the spatial prediction. A prediction sample may mean a sample within a prediction block.

The inter prediction unit 110 may include a motion estimation unit and a motion compensation unit.

When the prediction mode is the inter mode, the motion prediction unit may search for a region that best matches the target block from the reference image in the motion prediction process, and derive motion vectors for the target block and the searched region using the searched region. can do. At this time, the motion prediction unit may use a search area as a search target area.

The reference picture may be stored in the reference picture buffer 190, and when encoding and/or decoding of the reference picture is processed, the encoded and/or decoded reference picture may be stored in the reference picture buffer 190.

As a decoded picture is stored, the reference picture buffer 190 may be a decoded picture buffer (DPB).

The motion compensator may generate a prediction block for the target block by performing motion compensation using a motion vector. Here, the motion vector may be a 2D vector used for inter prediction. Also, the motion vector may indicate an offset between the target image and the reference image.

The motion estimation unit and the motion compensation unit may generate a prediction block by applying an interpolation filter to a partial region in a reference image when a motion vector has a non-integer value. In order to perform inter prediction or motion compensation, methods of motion prediction and motion compensation of PUs included in the CU based on the CU include skip mode, merge mode, and advanced motion vector prediction (Advanced Motion Vector Prediction (AMVP) mode and current picture reference mode may be determined, and inter prediction or motion compensation may be performed according to each mode.

The subtractor 125 may generate a residual block that is a difference between the target block and the prediction block. A residual block may also be referred to as a residual signal.

The residual signal may mean a difference between the original signal and the prediction signal. Alternatively, the residual signal may be a signal generated by transforming, quantizing, or transforming and quantizing the difference between the original signal and the predicted signal. A residual block may be a residual signal for a block unit.

The transform unit 130 may generate transform coefficients by performing transform on the residual block, and may output the generated transform coefficients. Here, the transform coefficient may be a coefficient value generated by performing transform on the residual block.

The conversion unit 130 may use one of a plurality of predefined conversion methods in performing the conversion.

A plurality of predefined transform methods may include discrete cosine transform (DCT), discrete sine transform (DST), and Karhunen-Loeve transform (KLT) based transform. there is.

A transform method used for transforming the residual block may be determined according to at least one of coding parameters for the target block and/or neighboring blocks. For example, the transform method may be determined based on at least one of an inter prediction mode for a PU, an intra prediction mode for a PU, a TU size, and a TU shape. Alternatively, transformation information indicating a transformation method may be signaled from the encoding apparatus 100 to the decoding apparatus 200.

When a transform skip mode is applied, the transform unit 130 may skip transforming the residual block.

A quantized transform coefficient level or a quantized level may be generated by applying quantization to the transform coefficients. Hereinafter, in the embodiments, a quantized transform coefficient level and a quantized level may also be referred to as a transform coefficient.

The quantization unit 140 may generate a quantized transform coefficient level (ie, a quantized level or a quantized coefficient) by quantizing a transform coefficient according to a quantization parameter. The quantization unit 140 may output the generated quantized transform coefficient level. At this time, the quantization unit 140 may quantize the transform coefficient using a quantization matrix.

The entropy encoding unit 150 may generate a bitstream by performing entropy encoding according to a probability distribution based on values calculated by the quantization unit 140 and/or coding parameter values calculated in the encoding process. . The entropy encoding unit 150 may output the generated bitstream.

The entropy encoding unit 150 may perform entropy encoding on information about pixels of an image and information for decoding an image. For example, information for decoding an image may include a syntax element and the like.

When entropy encoding is applied, a small number of bits may be allocated to a symbol having a high probability of occurrence, and a large number of bits may be allocated to a symbol having a low probability of occurrence. As symbols are represented through such allocation, the size of bitstrings for symbols that are encoding targets can be reduced. Therefore, compression performance of image encoding can be improved through entropy encoding.

In addition, the entropy encoding unit 150 uses exponential golomb, context-adaptive variable length coding (CAVLC), and context-adaptive binary arithmetic coding (Context-Adaptive Binary Coding) for entropy encoding. A coding method such as Arithmetic Coding (CABAC) may be used. For example, the entropy encoding unit 150 may perform entropy encoding using a variable length coding/code (VLC) table. For example, the entropy encoding unit 150 may derive a binarization method for a target symbol. Also, the entropy encoding unit 150 may derive a probability model of the target symbol/bin. The entropy encoding unit 150 may perform arithmetic encoding using the derived binarization method, probability model, and context model.

The entropy encoding unit 150 may change coefficients in the form of a 2-dimensional block into a form of a 1-dimensional vector through a transform coefficient scanning method in order to encode the quantized transform coefficient level.

A coding parameter may be information required for encoding and/or decoding. The coding parameter may include information that is encoded in the encoding device 100 and transmitted from the encoding device 100 to the decoding device, and may include information that can be derived in an encoding or decoding process. For example, as information transmitted to the decoding device, there is a syntax element.

A coding parameter may include information derived from an encoding process or a decoding process, as well as information (or flags and indexes, etc.) encoded in an encoding device and signaled from an encoding device to a decoding device, such as a syntax element. there is. Also, the coding parameter may include information required for encoding or decoding an image. For example, the size of the unit/block, the shape of the unit/block, the depth of the unit/block, division information of the unit/block, division structure of the unit/block, information indicating whether the unit/block is divided in the form of a quad tree, Information indicating whether the unit/block is split into a binary tree, the split direction of the binary tree (horizontal or vertical), the split type of the binary tree (symmetric or asymmetric), the unit/block is a ternary tree information indicating whether it is split into ternary tree type split direction (horizontal direction or vertical direction), ternary tree type split type (symmetric split or asymmetric split, etc.), unit/block multi-type tree information indicating whether or not to split into a multi-type tree, combination and direction (horizontal or vertical, etc.) of multi-type tree split, split type of multi-type tree split (symmetric split or asymmetric split), multi-type Split tree in tree form (binary tree or ternary tree), type of prediction mode (intra prediction or inter prediction), intra prediction mode/direction, intra luma prediction mode/direction, intra chroma prediction mode/direction, intra partitioning information, inter prediction Split information, coding block split flag, prediction block split flag, transform block split flag, reference sample filtering method, reference sample filter tap (tap), reference sample filter coefficient, prediction block filtering method, prediction block filter tap, prediction block filter coefficient , prediction block boundary filtering method, prediction block boundary filter tap, prediction block boundary filter coefficient, inter prediction mode, motion information, motion vector, motion vector difference, reference picture index, inter prediction direction, inter prediction indicator, prediction list utilization ) flag, reference picture list, reference picture, POC, motion vector predictor, motion vector prediction index, motion vector prediction candidate, motion vector candidate list, information indicating whether merge mode is used, merge index, merge candidate, merge candidate list , information indicating whether skip mode is used, interpolation filter type, interpolation filter filter tap, interpolation filter filter coefficient, motion vector magnitude, motion vector representation accuracy, transform type, transform size, and primary transform. Information indicating whether or not to use, information indicating whether to use additional (secondary) transform, primary transform selection information (or primary transform index), secondary transform selection information (or secondary transform index), residual Information indicating the presence or absence of a signal, a coded block pattern, a coded block flag, a quantization parameter, a residual quantization parameter, a quantization matrix, information about an intra-loop filter, and an intra-loop filter Information indicating whether to apply, coefficients of intra-loop filter, filter tap of intra-loop, shape/form of intra-loop filter, information indicating whether to apply deblocking filter, information indicating whether to apply deblocking filter, coefficient, filter tap of deblocking filter, strength of deblocking filter, shape/shape of deblocking filter, information indicating whether to apply adaptive sample offset, adaptive sample offset value, adaptive sample offset category, adaptive sample Offset type, information indicating whether the adaptive in-loop filter is applied, coefficients of the adaptive in-loop filter, filter tap of the adaptive in-loop filter, shape/shape of the adaptive in-loop filter Shape, binarization/inverse binarization method, context model, context model determination method, context model update method, information indicating whether regular mode is performed, information indicating whether bypass mode is performed, significant coefficient flags, last Significant coefficient flag, coefficient group unit coding flag, last significant coefficient position, flag indicating whether coefficient value is greater than 1, flag indicating whether coefficient value is greater than 2, flag indicating whether coefficient value is greater than 3 , remaining coefficient value information, sign information, reconstructed luma sample, reconstructed chroma sample, context bin, bypass bin, residual luma sample, residual chroma sample, transform coefficient, luma transform coefficient, chroma transform coefficient, Quantized level, luma quantized level, chroma quantized level, transform coefficient level, luma transform coefficient level, chroma transform coefficient level, transform coefficient level scanning method, size of motion vector search region from the side of a decoding device, decoding device The shape of the motion vector search area on the side, the number of motion vector searches on the side of the decoding device, CTU size, minimum block size, maximum block size, maximum block depth, minimum block depth, image display/output order, slice identification information , slice type, slice division information, tile group identification information, tile group type, tile group division information, tile identification information, tile type, tile division information, picture type, bit depth, input sample bit depth, reconstructed sample bit depth , residual sample bit depth, transform coefficient bit depth, quantized level bit depth, luma signal information, chroma signal information, a color space of a target block, and a value of at least one of a color space of a residual block, Combined forms or statistics can be included in the coding parameters. In addition, information related to the aforementioned coding parameters may also be included in the coding parameters. Information used to calculate and/or derive the aforementioned coding parameters may also be included in the coding parameters. Information calculated or derived using the aforementioned coding parameters may also be included in the coding parameters.

The primary transform selection information may indicate a primary transform applied to the target block.

The secondary transform selection information may indicate a secondary transform applied to the target block.

The residual signal may represent a difference between the original signal and the predicted signal. Alternatively, the residual signal may be a signal generated by transforming a difference between the original signal and the predicted signal. Alternatively, the residual signal may be a signal generated by transforming and quantizing a difference between the original signal and the predicted signal. A residual block may be a residual signal for a block.

Here, signaling information may mean that the encoding device 100 includes entropy-encoded information generated by performing entropy encoding on a flag or index in a bitstream, , in the decoding apparatus 200, may mean obtaining information by performing entropy decoding on entropy-encoded information extracted from a bitstream. Here, the information may include flags and indexes.

A signal may refer to signaled information. Hereinafter, information on images and blocks may be referred to as signals. Also, in the following, the terms “information” and “signal” may be used interchangeably and may be used interchangeably. For example, the specific signal may be a signal representing a specific block. An original signal may be a signal representing a target block. A prediction signal may be a signal representing a prediction block. A residual signal may be a signal representing a residual block.

A bitstream may include information according to a specified syntax. The encoding device 100 may generate a bitstream including information according to a specified syntax. The encoding device 200 may obtain information from a bitstream according to a specified syntax.

Since encoding through inter prediction is performed by the encoding apparatus 100, the encoded target image may be used as a reference image for other image(s) to be processed later. Accordingly, the encoding apparatus 100 may reconstruct or decode the encoded target image and store the reconstructed or decoded image in the reference picture buffer 190 as a reference image. Inverse quantization and inverse transformation may be performed on the encoded target image for decoding.

The quantized level may be inversely quantized in the inverse quantization unit 160 and inversely transformed in the inverse transformation unit 170 . The inverse quantization unit 160 may generate inverse quantized coefficients by performing inverse quantization on the quantized level. The inverse transform unit 170 may generate inverse quantized and inverse transformed coefficients by performing an inverse transform on the inverse quantized coefficients.

The inverse quantized and inverse transformed coefficients may be combined with the prediction block through the adder 175. A reconstructed block may be generated by adding the inverse quantized and inverse transformed coefficients and the prediction block. Here, the inverse quantized and/or inverse transformed coefficient may mean a coefficient on which at least one of dequantization and inverse-transformation has been performed, and may mean a reconstructed residual block. Here, the reconstructed block may mean a recovered block or a decoded block.

The reconstructed block may pass through the filter unit 180 . The filter unit 180 may include at least one of a deblocking filter, a sample adaptive offset (SAO), an adaptive loop filter (ALF), and a non-local filter (NLF). One or more may be applied to a reconstructed sample, reconstructed block or reconstructed picture. The filter unit 180 may also be referred to as an in-loop filter.

The deblocking filter may remove block distortion generated at a boundary between blocks in a reconstructed picture. In order to determine whether to apply the deblocking filter, it may be determined whether to apply the deblocking filter to the target block based on pixel(s) included in several columns or rows included in the block.

When a deblocking filter is applied to a target block, the applied filter may vary according to the required strength of deblocking filtering. In other words, among different filters, a filter determined according to the strength of deblocking filtering may be applied to the target block. When a deblocking filter is applied to the target block, a long-tap filter, a strong filter, a weak filter, and a Gaussian filter are selected according to the strength of the deblocking filtering required. ) may be applied to the target block.

Also, when vertical filtering and horizontal filtering are performed on a target block, horizontal filtering and vertical filtering may be processed in parallel.

SAO may add an appropriate offset to a pixel value of a pixel to compensate for a coding error. The SAO may perform correction using an offset on a difference between an original image and an image to which deblocking is applied in units of pixels for an image to which deblocking is applied. In order to perform offset correction on an image, a method of dividing pixels included in the image into a certain number of areas, determining an area to be offset from among the divided areas, and applying the offset to the determined area will be used. Alternatively, a method of applying an offset in consideration of edge information of each pixel of the image may be used.

ALF may perform filtering based on a value obtained by comparing the reconstructed image and the original image. After dividing the pixels included in the image into predetermined groups, a filter to be applied to each divided group may be determined, and filtering may be performed differentially for each group. Information related to whether to apply the adaptive loop filter may be signaled for each CU. This information may be signaled for the luma signal. The shape and filter coefficients of ALF to be applied to each block may be different for each block. Alternatively, a fixed type of ALF may be applied to the block regardless of the characteristics of the block.

The non-local filter may perform filtering based on reconstructed blocks similar to the target block. A region similar to the target block may be selected from the reconstructed image, and filtering of the target block may be performed using statistical properties of the selected similar region. Information related to whether to apply the non-local filter may be signaled to the CU. Also, shapes and filter coefficients of non-local filters to be applied to blocks may be different according to blocks.

A reconstructed block or a reconstructed image that has passed through the filter unit 180 may be stored in the reference picture buffer 190 as a reference picture. A reconstructed block that has passed through the filter unit 180 may be part of a reference picture. In other words, the reference picture may be a reconstructed picture composed of reconstructed blocks that have passed through the filter unit 180. The stored reference picture can then be used for inter prediction or motion compensation.

2 is a block diagram showing a configuration according to an embodiment of a decoding device to which the present invention is applied.

The decoding device 200 may be a decoder, a video decoding device, or an image decoding device.

Referring to FIG. 2 , the decoding apparatus 200 includes an entropy decoding unit 210, an inverse quantization unit 220, an inverse transform unit 230, an intra prediction unit 240, an inter prediction unit 250, and a switch 245. , an adder 255, a filter unit 260, and a reference picture buffer 270.

The decoding device 200 may receive the bitstream output from the encoding device 100. The decoding apparatus 200 may receive a bitstream stored in a computer readable recording medium or may receive a bitstream streamed through a wired/wireless transmission medium.

The decoding apparatus 200 may perform intra mode and/or inter mode decoding on a bitstream. Also, the decoding apparatus 200 may generate a reconstructed image or a decoded image through decoding, and output the generated reconstructed image or decoded image.

For example, conversion to an intra mode or an inter mode according to a prediction mode used for decoding may be performed by the switch 245 . When the prediction mode used for decoding is an intra mode, the switch 245 may be switched to intra mode. When the prediction mode used for decoding is the inter mode, the switch 245 may be switched to inter mode.

The decoding apparatus 200 may obtain a reconstructed residual block by decoding the input bitstream and generate a prediction block. When the reconstructed residual block and the prediction block are obtained, the decoding apparatus 200 may generate a reconstructed block to be decoded by summing the reconstructed residual block and the prediction block.

The entropy decoding unit 210 may generate symbols by performing entropy decoding on the bitstream based on the probability distribution of the bitstream. The generated symbols may include symbols in the form of quantized transform coefficient levels (ie, quantized levels or quantized coefficients). Here, the entropy decoding method may be similar to the above-described entropy encoding method. For example, the entropy decoding method may be a reverse process of the above-described entropy encoding method.

The entropy decoding unit 210 may change a coefficient in the form of a 1-dimensional vector into a form of a 2-dimensional block through a transform coefficient scanning method in order to decode the quantized transform coefficient level.

For example, the coefficients may be changed into a 2D block form by scanning the coefficients of a block using an upper-right diagonal scan. Alternatively, which scan to be used among the upper-right diagonal scan, vertical scan, and horizontal scan may be determined according to the size of the block and/or the intra prediction mode.

The quantized coefficient may be inversely quantized in the inverse quantization unit 220 . The inverse quantization unit 220 may generate inverse quantized coefficients by performing inverse quantization on the quantized coefficients. In addition, the inverse quantized coefficient may be inversely transformed in the inverse transform unit 230 . The inverse transform unit 230 may generate a reconstructed residual block by performing an inverse transform on the inverse quantized coefficient. As a result of performing inverse quantization and inverse transformation on the quantized coefficients, a reconstructed residual block may be generated. In this case, the inverse quantization unit 220 may apply a quantization matrix to the quantized coefficients in generating the reconstructed residual block.

When the intra mode is used, the intra predictor 240 may generate a prediction block by performing spatial prediction on the target block using pixel values of previously decoded blocks adjacent to the target block.

The inter prediction unit 250 may include a motion compensation unit. Alternatively, the inter prediction unit 250 may be referred to as a motion compensation unit.

When the inter mode is used, the motion compensator may generate a prediction block by performing motion compensation on the target block using a motion vector and a reference image stored in the reference picture buffer 270 .

When the motion vector has a non-integer value, the motion compensator may apply an interpolation filter to a partial region in the reference image and generate a prediction block using the reference image to which the interpolation filter is applied. The motion compensation unit may determine which mode among skip mode, merge mode, AMVP mode, and current picture reference mode is a motion compensation method used for a PU included in a CU based on the CU to perform motion compensation, and the determined mode Accordingly, motion compensation may be performed.

The reconstructed residual block and the prediction block may be added through an adder 255. Adder 255 may produce a reconstructed block by adding the reconstructed residual block and the prediction block.

The reconstructed block may pass through the filter unit 260 . The filter unit 260 may apply at least one of a deblocking filter, an SAO filter, an ALF filter, and a non-local filter to a reconstructed block or a reconstructed image. A reconstructed image may be a picture including a reconstructed block.

The filter unit 260 may output a reconstructed image.

The reconstructed block and/or the reconstructed image that has passed through the filter unit 260 may be stored as a reference picture in the reference picture buffer 270 . A reconstructed block that has passed through the filter unit 260 may be part of a reference picture. In other words, the reference picture may be a reconstructed image composed of reconstructed blocks that have passed through the filter unit 260 . The stored reference picture can then be used for inter prediction and/or motion compensation.

3 is a diagram schematically illustrating a division structure of an image when encoding and decoding an image.

3 may schematically show an example in which one unit is divided into a plurality of sub-units.

In order to efficiently segment an image, a coding unit (CU) may be used in encoding and decoding. A unit may be a term that collectively refers to 1) a block including image samples and 2) a syntax element. For example, “division of a unit” may mean “division of a block corresponding to a unit”.

A CU may be used as a base unit for image encoding and/or decoding. In addition, the CU may be used as a unit to which one selected mode of intra mode and inter mode is applied in video encoding and/or decoding. In other words, in video encoding and/or decoding, it may be determined which mode among the intra mode and the inter mode is to be applied to each CU.

Also, a CU may be a base unit in encoding and/or decoding of prediction, transform, quantization, inverse transform, inverse quantization, and transform coefficients.

Referring to FIG. 3 , an image 300 may be sequentially divided into units of largest coding units (LCUs). For each LCU, a partition structure may be determined. Here, LCU may be used as the same meaning as Coding Tree Unit (CTU).

Division of a unit may mean division of a block corresponding to the unit. The block division information may include depth information about the depth of a unit. Depth information may indicate the number and/or degree of division of a unit. One unit may be hierarchically divided into a plurality of sub-units with depth information based on a tree structure.

Each divided sub-unit may have depth information. Depth information may be information indicating the size of a CU. Depth information may be stored for each CU.

Each CU may have depth information. When a CU is split, CUs generated by splitting may have a depth increased by 1 from the depth of the split CU.

The division structure may refer to a distribution of CUs in the LCU 310 to efficiently encode an image. This distribution may be determined according to whether one CU is to be divided into a plurality of CUs. The number of divided CUs may be a positive integer greater than or equal to 2, including 2, 4, 8 and 16, and the like.

The horizontal and vertical sizes of the CU generated by division may be smaller than the horizontal and vertical sizes of the CU before division, depending on the number of CUs generated by division. For example, the horizontal size and vertical size of the CU generated by division may be half of the horizontal size and half of the vertical size of the CU before division.

A divided CU may be recursively divided into a plurality of CUs in the same way. By recursive division, at least one of the horizontal size and the vertical size of the divided CU may be reduced compared to at least one of the horizontal and vertical sizes of the CU before division.

The division of the CU may be made recursively to a predefined depth or to a predefined size.

For example, the depth of CU may have a value of 0 to 3. The size of the CU may range from 64x64 to 8x8 depending on the depth of the CU.

For example, the depth of the LCU 310 may be 0, and the depth of the smallest coding unit (SCU) may be a predefined maximum depth. Here, the LCU may be a CU having the largest coding unit size as described above, and the SCU may be a CU having the smallest coding unit size.

The division may start from the LCU 310, and the depth of the CU may increase by 1 whenever the horizontal size and/or the vertical size of the CU are reduced by the division.

For example, for each depth, a CU that is not split may have a size of 2Nx2N. In addition, in the case of a CU to be divided, a CU of 2Nx2N size may be divided into 4 CUs having a size of NxN. The size of N can be halved every time the depth increases by 1.

Referring to FIG. 3 , an LCU having a depth of 0 may be 64x64 pixels or a 64x64 block. 0 may be the minimum depth. A SCU with a depth of 3 can be 8x8 pixels or an 8x8 block. 3 may be the maximum depth. In this case, the CU of the 64x64 block, which is the LCU, may be expressed as a depth of 0. A CU of a 32x32 block can be represented with a depth of 1. A CU of a 16x16 block can be represented with a depth of 2. A CU of an 8x8 block, which is an SCU, can be expressed as a depth of 3.

Information on whether the CU is split may be expressed through split information of the CU. The division information may be 1 bit of information. All CUs except for the SCU may include partition information. For example, a value of partition information of a CU that is not split may be a first value, and a value of partition information of a CU that is split may be a second value. When the split information indicates whether the CU splits, the first value may be 0 and the second value may be 1.

For example, when one CU is divided into 4 CUs, the horizontal and vertical sizes of each CU of the 4 CUs generated by the division are half of the horizontal size and half of the vertical size of the CU before the division, respectively. can When a 32x32 CU is divided into 4 CUs, the sizes of the 4 CUs may be 16x16. When one CU is divided into 4 CUs, it can be said that the CU is divided in a quad-tree form. In other words, it can be seen that quad-tree partitioning is applied to the CU.

For example, when one CU is split into two CUs, the horizontal size or vertical size of each CU of the two CUs generated by the split is half of the horizontal size or half of the vertical size of the CU before the split, respectively. can When a CU of 32x32 size is vertically divided into two CUs, the sizes of the two divided CUs may be 16x32. When a CU of 32x32 size is horizontally divided into two CUs, the sizes of the two divided CUs may be 32x16. When one CU is divided into two CUs, it can be said that the CU is divided in a binary-tree form. In other words, it can be seen that binary-tree partitioning has been applied to the CU.

For example, when one CU is divided into three CUs, the three divided CUs may be generated by dividing the horizontal or vertical size of the CU before being divided at a ratio of 1:2:1. For example, when a CU of 16x32 size is divided into 3 CUs in the horizontal direction, the 3 divided CUs may have sizes of 16x8, 16x16 and 16x8, respectively, from the top. For example, when a CU of size 32x32 is divided into three CUs in the vertical direction, the three divided CUs may have sizes of 8x32, 16x32, and 8x32 from the left, respectively. When one CU is divided into three CUs, it can be said that the CU is divided in a ternary-tree form. In other words, it can be seen that ternary-tree partitioning is applied to the CU.

Both quad-tree partitioning and binary-tree partitioning are applied to the LCU 310 of FIG. 3 .

In the encoding apparatus 100, a 64x64 Coding Tree Unit (CTU) may be divided into a plurality of smaller CUs by a recursive quad-tree structure. One CU can be divided into 4 CUs with equal sizes. CUs can be partitioned recursively, and each CU can have a structure of a quad tree.

Through recursive segmentation on the CU, an optimal segmentation method that results in the smallest rate-distortion ratio can be selected.

The CTU 320 of FIG. 3 is an example of a CTU to which quad tree partitioning, binary tree partitioning, and ternary tree partitioning are all applied.

As described above, in order to divide the CTU, at least one of quad tree partitioning, binary tree partitioning, and ternary tree partitioning may be applied to the CTU. Partitions can be applied based on a specified priority order.

For example, quad tree partitioning may be preferentially applied to CTUs. A CU that cannot be further divided into a quad tree may correspond to a leaf node of a quad tree. A CU corresponding to a leaf node of a quad tree may be a root node of a binary tree and/or a ternary tree. That is, a CU corresponding to a leaf node of a quad tree may be split in the form of a binary tree or a ternary tree, or may not be split any further. At this time, quad tree splitting is not applied again to a CU generated by applying binary tree splitting or ternary tree splitting to a CU corresponding to a leaf node of a quad tree, so that block splitting and/or signaling of block splitting information is performed. can be done effectively.

Splitting of a CU corresponding to each node of the quad tree may be signaled using quad splitting information. Quad splitting information having a first value (eg, “1”) may indicate that the CU is split in the form of a quad tree. Quad splitting information having a second value (eg, “0”) may indicate that the CU is not split in the form of a quad tree. The quad division information may be a flag having a specified length (eg, 1 bit).

There may be no priority between binary tree splitting and ternary tree splitting. That is, a CU corresponding to a leaf node of a quad tree may be partitioned in a binary tree form or a ternary tree form. In addition, a CU generated by binary tree splitting or ternary tree splitting may be split again into a binary tree shape or a ternary tree shape, or may not be split any further.

Partitioning in the case where there is no priority between binary tree partitioning and ternary tree partitioning may be referred to as multi-type tree partitioning. That is, a CU corresponding to a leaf node of a quad tree may become a root node of a multi-type tree. Splitting of a CU corresponding to each node of the multi-type tree may be signaled using at least one of information indicating whether or not the multi-type tree is split, splitting direction information, and splitting tree information. For splitting of a CU corresponding to each node of the multi-type tree, information indicating whether to split or not, splitting direction information, and splitting tree information may be signaled sequentially.

For example, information indicating whether a multi-type tree having a first value (eg, “1”) is split may indicate that the corresponding CU is split into a multi-type tree. Information indicating whether the multi-type tree having a second value (eg, “0”) is split may indicate that the corresponding CU is not split into a multi-type tree.

When a CU corresponding to each node of a multi-type tree is split into a multi-type tree, the corresponding CU may further include splitting direction information.

The splitting direction information may indicate a splitting direction of multi-type tree splitting. The division direction information having a first value (eg, “1”) may indicate that the corresponding CU is divided in the vertical direction. The division direction information having a second value (eg, “0”) may indicate that the corresponding CU is divided in the horizontal direction.

When a CU corresponding to each node of a multi-type tree is split into a multi-type tree, the corresponding CU may further include split tree information. Split tree information may indicate a tree used for multi-type tree split.

For example, split tree information having a first value (eg, “1”) may indicate that the corresponding CU is split in the form of a binary tree. Split tree information having a second value (eg, “0”) may indicate that the corresponding CU is split in the form of a ternary tree.

Here, each of the aforementioned information indicating whether to split or not, split tree information, and split direction information may be a flag having a specified length (eg, 1 bit).

At least one of the above-described quad splitting information, information indicating whether to split a multi-type tree, splitting direction information, and splitting tree information may be entropy-encoded and/or entropy-decoded. For entropy encoding/decoding of such information, information of neighboring CUs adjacent to the target CU may be used.

For example, it may be considered that there is a high probability that the division shape of the left CU and/or the upper CU (that is, whether or not to split, the split tree, and/or the split direction) and the split shape of the target CU are similar to each other. Accordingly, context information for entropy encoding and/or entropy decoding of information of a target CU may be derived based on information of a neighboring CU. In this case, the information of the neighboring CU may include at least one of 1) quad split information, 2) information indicating whether the multi-type tree is split, 3) split direction information, and 4) split tree information of the neighboring CU.

As another embodiment, binary tree splitting may be preferentially performed among binary tree splitting and ternary tree splitting. That is, binary tree splitting is applied first, and a CU corresponding to a leaf node of the binary tree may be set as a root node of the ternary tree. In this case, quad tree splitting and binary tree splitting may not be performed on a CU corresponding to a node of a ternary tree.

A CU that is not further split by quad tree splitting, binary tree splitting, and/or ternary tree splitting may become a unit of encoding, prediction, and/or transformation. That is, for prediction and/or transformation, the CU may not be further split. Accordingly, a partitioning structure and partitioning information for dividing a CU into prediction units and/or transform units may not exist in the bitstream.

However, when the size of a CU serving as a unit of division is larger than the size of the maximum transform block, this CU may be recursively split until the size of the CU is less than or equal to the size of the maximum transform block. For example, when the size of a CU is 64x64 and the size of a maximum transform block is 32x32, the CU may be divided into four 32x32 blocks for transform. For example, when the size of a CU is 32x64 and the size of a maximum transform block is 32x32, the CU may be divided into two 32x32 blocks for transform.

In this case, information on whether the CU is split for conversion may not be separately signaled. Without signaling, whether to split the CU may be determined by comparing the horizontal size (and/or vertical size) of the CU and the horizontal size (and/or vertical size) of the largest transform block. For example, when the horizontal size of the CU is greater than the horizontal size of the largest transform block, the CU may be vertically divided into two parts. In addition, when the vertical size of the CU is greater than the vertical size of the largest transform block, the CU may be divided into two horizontally.

Information on the maximum size and/or minimum size of a CU and information on the maximum size and/or minimum size of a transform block may be signaled or determined at a higher level for a CU. For example, the upper level may be a sequence level, a picture level, a tile level, a tile group level, and a slice level. For example, the minimum size of a CU may be determined to be 4x4. For example, the maximum size of a transform block may be determined to be 64x64. For example, the minimum size of a transform block may be determined to be 4x4.

Information about the minimum size of a CU corresponding to a leaf node of a quad tree (ie quad tree minimum size) and/or the maximum depth of a path from the root node of a multi-type tree to a leaf node (ie multi-type tree maximum size). depth) may be signaled or determined at a higher level for the CU. For example, the higher level may be a sequence level, a picture level, a slice level, a tile group level, and a tile level. Information on the minimum size of the quad tree and/or information on the maximum depth of the multi-type tree may be separately signaled or determined for each intra-slice and inter-slice.

Difference information about the size of the CTU and the maximum size of the transform block may be signaled or determined at a higher level for the CU. For example, the higher level may be a sequence level, a picture level, a slice level, a tile group level, and a tile level. Information about the maximum size of the CU corresponding to each node of the binary tree (ie, the maximum size of the binary tree) may be determined based on the size and difference information of the CTU. The maximum size of the CU corresponding to each node of the ternary tree (ie, the maximum size of the ternary tree) may have a different value depending on the slice type. For example, within an intra slice, the maximum size of the ternary tree may be 32x32. Also, for example, within an inter slice, the maximum size of a ternary tree may be 128x128. For example, the minimum size of a CU corresponding to each node of a binary tree (say, the minimum size of a binary tree) and/or the minimum size of a CU corresponding to each node of a ternary tree (say, the minimum size of a ternary tree) is Can be set to a minimum size.

As another example, the maximum size of the binary tree and/or the maximum size of the ternary tree may be signaled or determined at the slice level. Also, the minimum size of the binary tree and/or the minimum size of the ternary tree may be signaled or determined at the slice level.

Based on the aforementioned various block sizes and various depths, quad splitting information, information indicating whether to split a multi-type tree, splitting tree information, and/or splitting direction information may or may not exist in the bitstream.

For example, if the size of the CU is not larger than the quad tree minimum size, the CU may not include quad splitting information, and the quad splitting information for the CU may be inferred as a second value.

For example, if the size (horizontal size and vertical size) of a CU corresponding to a node of a multi-type tree is larger than the binary tree maximum size (horizontal size and vertical size) and/or the ternary tree maximum size (horizontal size and vertical size) If larger, the CU may not be partitioned into binary tree form and/or ternary tree form. According to this determination method, information indicating whether to split the multi-type tree may not be signaled and may be inferred as a second value.

Alternatively, the size (horizontal size and vertical size) of a CU corresponding to a node of a multi-type tree is equal to the minimum size (horizontal size and vertical size) of a binary tree, or the size of a CU (horizontal size and vertical size) is equal to the size of a ternary tree. When equal to twice the minimum size (horizontal size and vertical size), the CU may not be split into a binary tree shape and/or a ternary tree shape. According to this determination method, information indicating whether to split the multi-type tree may not be signaled and may be inferred as a second value. This is because when the CU is divided into a binary tree shape and/or a ternary tree shape, a CU smaller than the minimum size of the binary tree and/or the minimum size of the ternary tree is generated.

Alternatively, binary tree splitting or ternary tree splitting may be limited based on the size of the virtual pipeline data unit (ie, pipeline buffer size). For example, binary tree splitting or ternary tree splitting may be limited when a CU is split into sub-CUs not suitable for a pipeline buffer size by binary tree splitting or ternary tree splitting. The pipeline buffer size may be equal to the size of the largest transform block (eg, 64X64).

For example, when the pipeline buffer size is 64X64, the following partitions can be limited.

- ternary tree split over NxM (N and/or M is 128) CUs

- horizontal binary tree partitioning over 128xN (N <= 64) CUs

- Vertical binary tree partitioning for Nx128 (N <= 64) CUs

Alternatively, when the depth in the multi-type tree of a CU corresponding to a node of the multi-type tree is equal to the maximum depth of the multi-type tree, the CU may not be split into a binary tree form and/or a ternary tree form. According to this determination method, information indicating whether to split the multi-type tree may not be signaled and may be inferred as a second value.

Alternatively, only when at least one of vertical binary tree splitting, horizontal binary tree splitting, vertical ternary tree splitting, and horizontal ternary tree splitting is possible for a CU corresponding to a node of a multi-type tree, a multi-type tree Information indicating whether to divide may be signaled. Otherwise, the CU may not be partitioned into binary tree form and/or ternary tree form. According to this determination method, information indicating whether to split the multi-type tree may not be signaled and may be inferred as a second value.

Alternatively, only when both vertical binary tree splitting and horizontal binary tree splitting are possible for a CU corresponding to a node of a multi-type tree, or both vertical ternary tree splitting and horizontal ternary tree splitting are possible, splitting direction information may be signaled. Otherwise, the division direction information may not be signaled and may be inferred as a value indicating a direction in which the CU may be divided.

Alternatively, only when both vertical binary tree splitting and vertical ternary tree splitting are possible for a CU corresponding to a node of a multi-type tree, or both horizontal binary tree splitting and horizontal ternary tree splitting are possible, split tree information may be signaled. Otherwise, the split tree information may not be signaled and may be inferred as a value indicating a tree applicable to split of a CU.

4 is a diagram illustrating a form of a prediction unit that a coding unit may include.

Among the CUs divided from the LCU, a CU that is not further divided may be divided into one or more prediction units (PUs).

A PU may be a basic unit for prediction. A PU may be coded and decoded in any one of skip mode, inter mode, and intra mode. A PU may be divided into various types according to each mode. For example, the target block described above with reference to FIG. 1 and the target block described above with reference to FIG. 2 may be PUs.

A CU may not be divided into PUs. When a CU is not divided into PUs, the size of a CU and the size of a PU may be the same.

In skip mode, there may not be a split within a CU. In the skip mode, a 2Nx2N mode 410 in which sizes of PU and CU are the same may be supported without division.

In inter mode, 8 divided types can be supported within the CU. For example, in inter mode, 2Nx2N mode 410, 2NxN mode 415, Nx2N mode 420, NxN mode 425, 2NxnU mode 430, 2NxnD mode 435, nLx2N mode 440 and nRx2N Mode 445 may be supported.

In intra mode, 2Nx2N mode 410 and NxN mode 425 may be supported.

In the 2Nx2N mode 410, a PU having a size of 2Nx2N may be encoded. A PU having a size of 2Nx2N may mean a PU having the same size as that of a CU. For example, a 2Nx2N PU may have a size of 64x64, 32x32, 16x16 or 8x8.

In the NxN mode 425, a PU having a size of NxN may be encoded.

For example, in intra prediction, when the size of a PU is 8x8, 4 divided PUs can be coded. The size of the divided PU may be 4x4.

When a PU is coded using an intra mode, the PU may be coded using one intra prediction mode among a plurality of intra prediction modes. For example, High Efficiency Video Coding (HEVC) technology may provide 35 intra prediction modes, and a PU may be coded in one intra prediction mode among the 35 intra prediction modes.

Which mode among the 2Nx2N mode 410 and the NxN mode 425 the PU will be encoded may be determined by a rate-distortion cost.

The encoding apparatus 100 may perform an encoding operation on a PU having a size of 2Nx2N. Here, the encoding operation may be encoding a PU in each of a plurality of intra prediction modes usable by the encoding apparatus 100 . An optimal intra prediction mode for a PU having a size of 2Nx2N may be derived through an encoding operation. An optimal intra prediction mode may be an intra prediction mode that generates a minimum rate-distortion cost for encoding a 2Nx2N PU among a plurality of intra prediction modes usable by the encoding apparatus 100 .

Also, the encoding apparatus 100 may sequentially perform an encoding operation on each PU of NxN-divided PUs. Here, the encoding operation may be encoding a PU in each of a plurality of intra prediction modes usable by the encoding apparatus 100 . An optimal intra prediction mode for an NxN-sized PU may be derived through an encoding operation. An optimal intra-prediction mode may be an intra-prediction mode that generates the lowest rate-distortion cost for encoding an NxN-sized PU among a plurality of intra-prediction modes usable by the encoding apparatus 100.

The encoding apparatus 100 may determine which of the 2Nx2N-sized PU and the NxN-sized PUs to encode based on the comparison of the rate-distortion cost of the 2Nx2N-sized PU and the rate-distortion costs of the NxN-sized PUs.

One CU may be divided into one or more PUs, and a PU may also be divided into a plurality of PUs.

For example, when one PU is divided into 4 PUs, the horizontal size and vertical size of each of the 4 PUs generated by the division are half of the horizontal size and half of the vertical size of the PU before the division, respectively. can When a 32x32 PU is divided into 4 PUs, the sizes of the 4 PUs may be 16x16. When one PU is divided into 4 PUs, it can be said that the PU is divided in a quad-tree shape.

For example, when one PU is split into two PUs, the horizontal size or vertical size of each PU of the two PUs generated by the split is half of the horizontal size or half of the vertical size of the PU before splitting, respectively. can When a 32x32 PU is vertically divided into 2 PUs, the sizes of the 2 PUs may be 16x32. When a 32x32-sized PU is horizontally divided into two PUs, the sizes of the two divided PUs may be 32x16. When one PU is divided into two PUs, it can be said that the PU is divided in the form of a binary-tree.

5 is a diagram illustrating a form of a transform unit that may be included in a coding unit.

A transform unit (TU) may be a basic unit used for transformation, quantization, inverse transformation, inverse quantization, entropy encoding, and entropy decoding processes within the CU.

A TU may have a square shape or a rectangular shape. The shape of the TU may be determined depending on the size and/or shape of the CU.

Among the CUs split from the LCU, a CU that is no longer split into CUs may be split into one or more TUs. In this case, the division structure of the TU may be a quad-tree structure. For example, as shown in FIG. 5 , one CU 510 may be divided one or more times according to a quad-tree structure. Through division, one CU 510 may be composed of TUs of various sizes.

When one CU is split two or more times, the CU can be considered to be split recursively. Through splitting, one CU may be composed of TUs having various sizes.

Alternatively, one CU may be divided into one or more TUs based on the number of vertical and/or horizontal lines dividing the CU.

A CU may be divided into symmetric TUs or asymmetric TUs. For division into asymmetric TUs, information on the size and/or shape of a TU may be signaled from the encoding device 100 to the decoding device 200. Alternatively, the size and/or shape of the TU may be derived from information on the size and/or shape of the CU.

A CU may not be divided into TUs. When a CU is not divided into TUs, the size of a CU and the size of a TU may be the same.

One CU may be divided into one or more TUs, and a TU may also be divided into a plurality of TUs.

For example, when one TU is divided into 4 TUs, the horizontal and vertical sizes of each TU of the 4 TUs generated by the division are half of the horizontal size and half of the vertical size of the TU before the division, respectively. can When a TU of 32x32 size is divided into 4 TUs, the sizes of the 4 divided TUs may be 16x16. When one TU is divided into 4 TUs, it can be said that the TU is divided in the form of a quad-tree.

For example, when one TU is split into two TUs, the horizontal size or vertical size of each TU of the two TUs generated by the split is half of the horizontal size or half of the vertical size of the TU before the split, respectively. can When a 32x32 size TU is vertically split into two TUs, the sizes of the two split TUs may be 16x32. When a TU of size 32x32 is horizontally divided into two TUs, the sizes of the two divided TUs may be 32x16. When one TU is split into two TUs, it can be said that the TU is split in a binary-tree format.

CUs may be divided in other ways than shown in FIG. 5 .

For example, one CU may be split into 3 CUs. The horizontal size or vertical size of the three divided CUs may be 1/4, 1/2, and 1/4 of the horizontal or vertical size of the CU before division, respectively.

For example, when a CU of 32x32 size is vertically divided into 3 CUs, the sizes of the 3 divided CUs may be 8x32, 16x32 and 8x32, respectively. In this way, when one CU is split into three CUs, it can be considered that the CU is split in the form of a ternary tree.

One of the exemplified quad tree-type partitioning, binary tree-type partitioning, and ternary tree-type partitioning may be applied for partitioning of the CU, and a plurality of partitioning schemes may be combined together and used for partitioning of the CU. . In this case, a case in which a plurality of partitioning methods are combined and used may be referred to as a composite tree type of partitioning.

6 shows division of a block according to an example.

In the process of encoding and/or decoding an image, a target block may be divided as shown in FIG. 6 . For example, the target block may be a CU.

For division of the target block, an indicator indicating division information may be signaled from the encoding apparatus 100 to the decoding apparatus 200. The division information may be information indicating how the target block is divided.

Splitting information includes a split flag (hereinafter referred to as “split_flag”), a quad-binary flag (hereinafter referred to as “QB_flag”), a quad-tree flag (hereinafter referred to as “quadtree_flag”), and a binary tree flag (hereinafter referred to as “binarytree_flag”). ") and a binary type flag (hereinafter referred to as "Btype_flag").

split_flag may be a flag indicating whether a block is split. For example, a value of 1 of split_flag may indicate that a block is split. A value of 0 of split_flag may indicate that the block is not split.

QB_flag may be a flag indicating in which form a block is divided into a quad tree form and a binary tree form. For example, a value of 0 of QB_flag may indicate that a block is divided in a quad tree form. A value of 1 of QB_flag may indicate that a block is divided in the form of a binary tree. Alternatively, a value of 0 of QB_flag may indicate that the block is divided in the form of a binary tree. A value of 1 for QB_flag may indicate that a block is divided in a quad tree form.

quadtree_flag may be a flag indicating whether a block is split into a quad tree. For example, a value of 1 of quadtree_flag may indicate that a block is split into a quad tree. A value of 0 of quadtree_flag may indicate that a block is not split in the form of a quad tree.

binarytree_flag may be a flag indicating whether a block is split in the form of a binary tree. For example, a value of 1 in binarytree_flag may indicate that a block is split into a binary tree. A value of 0 in binarytree_flag may indicate that a block is not split in the form of a binary tree.

Btype_flag may be a flag indicating whether the block is split vertically or horizontally when the block is split in the form of a binary tree. For example, a value of 0 in Btype_flag may indicate that a block is divided in a horizontal direction. A value of 1 of Btype_flag may indicate that a block is divided in a vertical direction. Alternatively, a value of 0 of Btype_flag may indicate that the block is divided in the vertical direction. A value of 1 of Btype_flag may indicate that the block is divided in the horizontal direction.

For example, partition information for the block of FIG. 6 can be derived by signaling at least one of quadtree_flag, binarytree_flag, and Btype_flag as shown in Table 1 below.

[Table 1]

For example, the split information for the block of FIG. 6 can be derived by signaling at least one of split_flag, QB_flag, and Btype_flag as shown in Table 2 below.

[Table 2]

The partitioning method may be limited to a quad tree, or a binary tree, depending on the size and/or shape of the block. When this restriction is applied, split_flag may be a flag indicating whether to split into a quad tree type or a flag indicating whether to split into a binary tree type. The size and shape of the block may be derived according to the depth information of the block, and the depth information may be signaled from the encoding apparatus 100 to the decoding apparatus 200.

When the size of a block falls within a specified range, only quad-tree partitioning may be possible. For example, the specified range may be defined by at least one of a maximum block size and a minimum block size in which quad-tree partitioning is possible.

Information indicating the maximum block size and/or minimum block size in which only quad tree-type division is possible may be signaled from the encoding apparatus 100 to the decoding apparatus 200 through a bitstream. Also, this information may be signaled for at least one unit of video, sequence, picture, parameter, tile group, and slice (or segment).

Alternatively, the maximum block size and/or the minimum block size may be fixed sizes predefined in the encoding apparatus 100 and the decoding apparatus 200. For example, when the size of a block is larger than 64x64 and smaller than 256x256, only quad-tree partitioning may be possible. In this case, split_flag may be a flag indicating whether to split into a quad tree.

If the size of the block is larger than the maximum transform block size, only a quad tree type of splitting may be possible. In this case, the divided block may be at least one of CU and TU.

In this case, split_flag may be a flag indicating whether to split into a quad tree.

When the size of a block falls within a specified range, only binary tree or ternary tree partitioning may be possible. Here, for example, the specified range may be defined by at least one of a maximum block size and a minimum block size in which binary tree or ternary tree partitioning is possible.

Information representing a maximum block size and/or a minimum block size in which only binary tree-type division or ternary tree-type division is possible may be signaled from the encoding apparatus 100 to the decoding apparatus 200 through a bitstream. In addition, this information may be signaled for at least one unit of sequence, picture, and slice (or segment).

Alternatively, the maximum block size and/or the minimum block size may be fixed sizes predefined in the encoding apparatus 100 and the decoding apparatus 200. For example, when the size of a block is greater than 8x8 and less than 16x16, only binary tree partitioning may be possible. In this case, split_flag may be a flag indicating whether to split into a binary tree type or a ternary tree type.

The above-described quad tree partitioning may be equally applied to binary tree and/or ternary tree partitioning.

The division of a block may be limited by the previous division. For example, when a block is divided into a specified binary tree shape to generate a plurality of divided blocks, each divided block may be additionally divided only into the specified tree shape. Here, the specified tree type may be at least one of a binary tree type, a ternary tree type, and a quad tree type.

When the horizontal size or vertical size of a divided block corresponds to a size that cannot be further divided, the aforementioned indicator may not be signaled.

7 is a diagram for explaining an embodiment of an intra prediction process.

Arrows extending from the center to the periphery of the graph of FIG. 7 may represent prediction directions of directional intra prediction modes. Also, numbers displayed adjacent to arrows may represent examples of mode values assigned to an intra prediction mode or a prediction direction of the intra prediction mode.

In FIG. 7 , the number 0 may indicate a planar mode, which is a non-directional intra prediction mode. Number 1 may indicate a DC mode, which is a non-directional intra prediction mode.

Intra coding and/or decoding may be performed using reference samples of neighboring blocks of the target block. A neighboring block may be a reconstructed neighboring block. A reference sample may mean a neighboring sample.

For example, intra encoding and/or decoding may be performed using a value of a reference sample included in a reconstructed neighboring block or a coding parameter.

The encoding apparatus 100 and/or the decoding apparatus 200 may generate a prediction block by performing intra prediction on the target block based on sample information in the target image. When performing intra prediction, the encoding apparatus 100 and/or the decoding apparatus 200 may generate a prediction block for a target block by performing intra prediction based on sample information in a target image. When intra prediction is performed, the encoding device 100 and/or the decoding device 200 may perform directional prediction and/or non-directional prediction based on at least one reconstructed reference sample.

A prediction block may refer to a block generated as a result of performing intra prediction. A prediction block may correspond to at least one of CU, PU, and TU.

A unit of a prediction block may be the size of at least one of CU, PU, and TU. The prediction block may have a square shape with a size of 2Nx2N or NxN. The size of NxN may include 4x4, 8x8, 16x16, 32x32, and 64x64.

Alternatively, the prediction block may be a square block having a size of 2x2, 4x4, 8x8, 16x16, 32x32, or 64x64, or may be a rectangular block having a size of 2x8, 4x8, 2x16, 4x16, or 8x16. there is.

Intra prediction may be performed according to an intra prediction mode for a target block. The number of intra prediction modes that the target block may have may be a predefined fixed value or may be a value determined differently according to properties of the prediction block. For example, the properties of the prediction block may include the size of the prediction block and the type of the prediction block. Also, the properties of the prediction block may indicate coding parameters for the prediction block.

For example, the number of intra prediction modes may be fixed to N regardless of the size of a prediction block. Alternatively, for example, the number of intra prediction modes may be 3, 5, 9, 17, 34, 35, 36, 65, 67, or 95.

The intra prediction mode may be a non-directional mode or a directional mode.

For example, the intra prediction mode may include two non-directional modes and 65 directional modes, corresponding to numbers 0 to 66 shown in FIG. 7 .

For example, when a specified intra prediction method is used, the intra prediction mode may include two non-directional modes and 93 directional modes, corresponding to numbers -14 to 80 shown in FIG. 7 .

The two non-directional modes may include DC mode and Planar mode.

The directional mode may be a prediction mode having a specific direction or a specific angle. The directional mode may also be referred to as an argular mode.

The intra prediction mode may be expressed as at least one of a mode number, a mode value, a mode angle, and a mode direction. That is to say, the terms “(mode) number of an intra prediction mode”, “(mode) value of an intra prediction mode”, “(mode) angle of an intra prediction mode” and “(mode) direction of an intra prediction mode) have the same meaning. and can be used interchangeably.

The number of intra prediction modes may be M. M may be 1 or more. In other words, the number of intra prediction modes may be M including the number of non-directional modes and the number of directional modes.

The number of intra prediction modes may be fixed to M regardless of block sizes and/or color components. For example, the number of intra prediction modes may be fixed to either 35 or 67 regardless of the size of a block.

Alternatively, the number of intra prediction modes may be different according to the shape, size, and/or color component type of a block.

For example, in FIG. 7 , directional prediction modes indicated by dotted lines may be applied only to prediction for non-square blocks.

For example, as the block size increases, the number of intra prediction modes may increase. Alternatively, as the block size increases, the number of intra prediction modes may decrease. When the size of a block is 4x4 or 8x8, the number of intra prediction modes may be 67. When the size of a block is 16x16, the number of intra prediction modes may be 35. When the size of a block is 32x32, the number of intra prediction modes may be 19. When the size of a block is 64x64, the number of intra prediction modes may be 7.

For example, the number of intra prediction modes may be different depending on whether the color component is a luma signal or a chroma signal. Alternatively, the number of intra prediction modes of the luma component block may be greater than the number of intra prediction modes of the chroma component block.

For example, in the case of a vertical mode having a mode value of 50, prediction may be performed in a vertical direction based on a pixel value of a reference sample. For example, in the case of a horizontal mode having a mode value of 18, prediction may be performed in a horizontal direction based on a pixel value of a reference sample.

Even in a directional mode other than the aforementioned mode, the encoding apparatus 100 and the decoding apparatus 200 may perform intra prediction on a target unit using a reference sample according to an angle corresponding to the directional mode.

An intra prediction mode positioned to the right of the vertical mode may be named a vertical-right mode. An intra-prediction mode positioned below the horizontal mode may be named a horizontal-below mode. For example, in FIG. 7 , intra prediction modes having a mode value of one of 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, and 66 are vertical It can be the right modes. Intra prediction modes having a mode value of 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, and 17 may be horizontal bottom modes.

The non-directional mode may include a DC mode and a planar mode. For example, the mode value of the DC mode may be 1. The mode value of the planner mode may be 0.

The directional mode may include an angular mode. Modes other than the DC mode and the planner mode among the plurality of intra prediction modes may be directional modes.

When the intra prediction mode is the DC mode, a prediction block may be generated based on an average of pixel values of a plurality of reference samples. For example, a pixel value of a prediction block may be determined based on an average of pixel values of a plurality of reference samples.

The number of intra prediction modes and the mode value of each intra prediction mode described above may be merely illustrative. The number of the aforementioned intra prediction modes and the mode value of each intra prediction mode may be differently defined according to embodiments, implementations, and/or needs.

In order to perform intra prediction on the target block, a step of checking whether samples included in the reconstructed neighboring block can be used as reference samples of the target block may be performed. If there is a sample that cannot be used as a reference sample of the target block among the samples of the neighboring block, a value generated by copying and/or interpolation using at least one sample value among the samples included in the reconstructed neighboring block This can be replaced with the sample value of a sample that cannot be used as a reference sample. If a value generated by copying and/or interpolation is replaced with a sample value of a sample, the sample may be used as a reference sample of the target block.

When intra prediction is used, a filter may be applied to at least one of a reference sample or a prediction sample based on at least one of an intra prediction mode and a size of a target block.

The type of filter applied to at least one of the reference sample and the prediction sample may differ according to at least one of an intra prediction mode of the target block, a size of the target block, and a shape of the target block. The type of filter may be classified according to one or more of a length of a filter tap, a value of a filter coefficient, and a filter strength. The length of the filter tap may mean the number of filter taps. Also, the number of filter taps may mean the length of a filter.

When the intra prediction mode is the planner mode, in generating the prediction block of the target block, the upper reference sample of the target sample, the left reference sample of the target sample, and the upper right reference sample of the target block according to the location of the prediction target sample in the prediction block. A sample value of the prediction target sample may be generated using a weight-sum of the lower left reference sample of the target block.

When the intra prediction mode is the DC mode, an average value of upper reference samples and left reference samples of the target block may be used to generate a prediction block of the target block. In addition, filtering using values of reference samples may be performed on specified rows or specified columns within the target block. The specified rows may be one or more top rows adjacent to the reference sample. The specified columns may be one or more left columns adjacent to the reference sample.

When the intra prediction mode is a directional mode, a prediction block may be generated using an upper reference sample, a left reference sample, an upper right reference sample, and/or a lower left reference sample of the target block.

Interpolation in units of real numbers may be performed to generate the prediction samples described above.

The intra prediction mode of the target block may be predicted from the intra prediction modes of neighboring blocks of the target block, and information used for prediction may be entropy encoded/decoded.

For example, if the intra prediction modes of the target block and the neighboring block are the same, it may be signaled that the intra prediction modes of the target block and the neighboring block are the same using a predefined flag.

For example, an indicator indicating an intra prediction mode identical to that of a target block among intra prediction modes of a plurality of neighboring blocks may be signaled.

If intra prediction modes of the target block and neighboring blocks are different from each other, information of the intra prediction mode of the target block may be encoded and/or decoded using entropy encoding and/or decoding.

8 is a diagram for explaining reference samples used in an intra prediction process.

The reconstructed reference samples used for intra prediction of the target block include below-left reference samples, left reference samples, above-left corner reference samples, and above reference samples. s and above-right reference samples, etc.

For example, the left reference samples may refer to reconstructed reference pixels adjacent to the left side of the target block. Top reference samples may refer to reconstructed reference pixels adjacent to the top of the target block. The top left corner reference sample may refer to a reconstructed reference pixel located at the top left corner of the target block. In addition, the lower left reference samples may refer to a reference sample located at the lower end of the left sample line among samples located on the same line as the left sample line composed of the left reference samples. The upper right reference samples may refer to reference samples located to the right of the upper pixel line among samples located on the same line as the upper sample line composed of the upper reference samples.

When the size of the target block is NxN, each of the lower left reference samples, the left reference samples, the upper reference samples, and the upper right reference samples may be N.

A prediction block may be generated through intra prediction of the target block. Generation of the prediction block may include determining values of pixels of the prediction block. The size of the target block and the prediction block may be the same.

Reference samples used for intra prediction of the target block may vary according to the intra prediction mode of the target block. A direction of an intra prediction mode may represent a dependency relationship between reference samples and pixels of a prediction block. For example, a value of a specified reference sample may be used as a value of one or more specified pixels of a prediction block. In this case, the specified reference sample and the specified one or more pixels of the prediction block may be samples and pixels specified as a straight line in the direction of the intra prediction mode. In other words, the value of the specified reference sample may be copied to a value of a pixel located in a direction opposite to the direction of the intra prediction mode. Alternatively, the value of a pixel of the prediction block may be a value of a reference sample located in the direction of the intra prediction mode based on the position of the pixel.

For example, when the intra prediction mode of the target block is a vertical mode, upper reference samples may be used for intra prediction. When the intra prediction mode is a vertical mode, a value of a pixel of a prediction block may be a value of a reference sample positioned vertically above the position of the pixel. Accordingly, the top reference samples that are top adjacent to the target block may be used for intra prediction. Also, values of pixels in one row of the prediction block may be the same as values of upper reference samples.

For example, when the intra prediction mode of the target block is a horizontal mode, left reference samples may be used for intra prediction. When the intra prediction mode is a horizontal mode, a value of a pixel of a prediction block may be a value of a reference sample located horizontally to the left of the pixel. Accordingly, left reference samples adjacent to the left of the target block may be used for intra prediction. Also, values of pixels of one column of the prediction block may be the same as values of left reference samples.

For example, when the mode value of the intra prediction mode of the target block is 34, at least some of the left reference samples, the upper left corner reference sample, and the upper reference samples may be used for intra prediction. When the mode value of the intra prediction mode is 34, the value of the pixel of the prediction block may be the value of the reference sample located diagonally at the top left of the pixel.

Also, when an intra prediction mode having a mode value of one of 52 to 66 is used, at least some of the upper right reference samples may be used for intra prediction.

Also, when an intra prediction mode having a mode value of one of 2 to 17 is used, at least some of lower left reference samples may be used for intra prediction.

Also, when an intra prediction mode having a mode value of one of 19 to 49 is used, the upper left corner reference sample may be used for intra prediction.

The number of reference samples used to determine the pixel value of one pixel of the prediction block may be one or two or more.

As described above, a pixel value of a pixel of a prediction block may be determined according to the position of the pixel and the position of the reference sample indicated by the direction of the intra prediction mode. When the position of the reference sample indicated by the position of the pixel and the direction of the intra prediction mode is an integer position, a value of one reference sample indicated by the integer position may be used to determine a pixel value of a pixel of the prediction block.

When the position of the reference sample indicated by the position of the pixel and the direction of the intra prediction mode is not an integer position, an interpolated reference sample may be generated based on two reference samples closest to the position of the reference sample. there is. The value of the interpolated reference sample may be used to determine a pixel value of a pixel of the predictive block. In other words, when the location of a pixel of the prediction block and the location of a reference sample indicated by the direction of the intra prediction mode indicate a gap between two reference samples, an interpolated value will be generated based on the values of the two samples. can

A prediction block generated by prediction may not be the same as the original target block. In other words, a prediction error, which is a difference between a target block and a prediction block, may exist, and a prediction error may also exist between pixels of the target block and pixels of the prediction block.

Hereinafter, the terms “difference”, “error” and “residual” may be used interchangeably and interchangeably.

For example, in the case of directional intra prediction, a larger prediction error may occur as the distance between a pixel of a prediction block and a reference sample increases. A discontinuity may occur between a prediction block generated by such a prediction error and a neighboring block.

Filtering of prediction blocks may be used to reduce prediction errors. Filtering may be adaptively applying a filter to a region considered to have a large prediction error among prediction blocks. For example, a region considered to have a large prediction error may be a boundary of a prediction block. Also, depending on the intra-prediction mode, a region considered to have a large prediction error among prediction blocks may be different, and filter characteristics may be different.

As shown in FIG. 8 , at least one of reference lines 0 to 3 may be used for intra prediction of a target block.

Each reference line in FIG. 8 may represent a reference sample line including one or more reference samples. The smaller the reference line number, the closer the reference sample line may be to the target block.

Samples of segment A and segment F may be obtained through padding using the nearest samples of segment B and segment E, respectively, instead of being obtained from reconstructed neighboring blocks.

Index information indicating a reference sample line to be used for intra prediction of a target block may be signaled. The index information may indicate a reference sample line used for intra prediction of a target block among a plurality of reference sample lines. For example, index information may have a value of 0 to 3.

When the upper boundary of the target block is the boundary of the CTU, only reference sample line 0 may be available. Therefore, in this case, index information may not be signaled. When a reference sample line other than reference sample line 0 is used, filtering of a prediction block described later may not be performed.

In the case of inter-color intra prediction, a prediction block for a target block of a second color component may be generated based on a corresponding reconstructed block of a first color component.

For example, the first color component may be a luma component, and the second color component may be a chroma component.

For intra prediction between color components, parameters of a linear model between a first color component and a second color component may be derived based on a template.

The template may include the top reference sample and/or the left reference sample of the target block, and the top reference sample and/or left reference sample of the reconstructed block of the first color component corresponding to these reference samples. there is.

For example, the parameters of the linear model are: 1) the value of the sample of the first color component having the largest value among the samples in the template, 2) the value of the sample of the second color component corresponding to the sample of this first color component, 3) the value of the sample of the first color component having the minimum value among the samples in the template and 4) the value of the sample of the second color component corresponding to the sample of the first color component.

When parameters of the linear model are derived, a prediction block for the target block may be generated by applying the corresponding reconstructed block to the linear model.

Depending on the image format, sub-sampling may be performed on the neighboring samples of the reconstructed block of the first color component and the corresponding reconstructed block. For example, when one sample of the second color component corresponds to four samples of the first color component, one corresponding sample may be calculated by subsampling the four samples of the first color component. there is. When subsampling is performed, derivation of parameters of the linear model and intra prediction between color components may be performed based on the subsampled corresponding samples.

Whether intra prediction between color components is performed and/or a template range may be signaled as an intra prediction mode.

A target block may be divided into 2 or 4 sub-blocks in a horizontal direction and/or a vertical direction.

Divided sub-blocks may be sequentially reconstructed. That is, as intra prediction is performed on a sub-block, a sub-prediction block for the sub-block may be generated. In addition, as inverse quantization and/or inverse transformation are performed on the sub-block, a sub-residual block for the sub-block may be generated. A reconstructed sub-block may be generated by adding the sub-prediction block to the sub-residual block. The reconstructed subblock may be used as a reference sample for intra prediction of a subblock of a later order.

A sub-block may be a block including a specified number (eg, 16) or more samples. Thus, for example, when the target block is an 8x4 block or a 4x8 block, the target block may be divided into two sub-blocks. Also, if the target block is a 4x4 block, the target block cannot be divided into sub blocks. If the target block has other sizes, the target block may be divided into 4 sub-blocks.

Information about whether intra prediction based on these sub-blocks is performed and/or a division direction (horizontal direction or vertical direction) may be signaled.

Such sub-block-based intra prediction may be limited to being performed only when the reference sample line 0 is used. When sub-block-based intra prediction is performed, filtering of a prediction block described later may not be performed.

A final prediction block may be generated by performing filtering on a prediction block generated by intra prediction.

Filtering may be performed by applying specific weights to the filtering target sample, the left reference sample, the top reference sample, and/or the top left reference sample.

A weight and/or a reference sample (or a range of reference samples or a location of the reference sample) used for filtering may be determined based on at least one of a block size, an intra prediction mode, and a position of a sample to be filtered within a prediction block. there is.

For example, filtering may be performed only for a specified intra prediction mode (eg, DC mode, planar mode, vertical mode, horizontal mode, diagonal mode, and/or adjacent diagonal mode).

The adjacent diagonal mode may be a mode having a number obtained by adding k to the number of the diagonal mode, or a mode having a number obtained by subtracting k from the number of the diagonal mode. In other words, the number of adjacent diagonal modes may be the sum of the number of diagonal modes and k, and may be the difference between the number of diagonal modes and k. For example, k may be a positive integer of 8 or less.

The intra prediction mode of the target block may be derived using intra prediction modes of neighboring blocks existing around the target block, and the derived intra prediction mode may be entropy-encoded and/or entropy-decoded.

For example, if the intra-prediction mode of the target block and the intra-prediction mode of the neighboring block are the same, information that the intra-prediction mode of the target block and the intra-prediction mode of the neighboring block are the same may be signaled using specified flag information. .

Also, for example, indicator information about a neighboring block having the same intra prediction mode as that of the target block among intra prediction modes of a plurality of neighboring blocks may be signaled.

For example, if the intra prediction mode of the target block and the intra prediction mode of the neighboring block are different, entropy encoding and/or entropy decoding based on the intra prediction mode of the neighboring block is performed to obtain information about the intra prediction mode of the target block. Entropy encoding and/or entropy decoding may be performed.

9 is a diagram for explaining an embodiment of an inter prediction process.

The rectangle shown in FIG. 9 may represent an image (or picture). Also, arrows in FIG. 9 may indicate prediction directions. An arrow pointing from a first picture to a second picture may indicate that the second picture refers to the first picture. That is, an image may be encoded and/or decoded according to a prediction direction.

Each image may be classified into an intra picture (I picture), a uni-prediction picture (P picture), and a bi-prediction picture (B picture) according to an encoding type. Each picture may be coded and/or decoded according to the coding type of each picture.

When a target image to be encoded is an I-picture, the target image may be encoded using data within the image itself without inter prediction referring to another image. For example, an I picture can be coded only with intra prediction.

When the target image is a P picture, the target image may be coded through inter prediction using only reference pictures existing in one direction. Here, the unidirectional direction may be a forward direction or a reverse direction.

When the target video is a B picture, the target video may be coded through inter prediction using reference pictures existing in both directions or inter prediction using reference pictures existing in one of forward and backward directions. Here, both directions may be forward and reverse directions.

P-pictures and B-pictures that are coded and/or decoded using reference pictures may be regarded as pictures in which inter prediction is used.

Below, inter prediction in inter mode according to an embodiment is described in detail.

Inter prediction or motion compensation may be performed using a reference image and motion information.

In the inter mode, the encoding apparatus 100 may perform inter prediction and/or motion compensation on a target block. The decoding apparatus 200 may perform inter prediction and/or motion compensation corresponding to inter prediction and/or motion compensation performed in the encoding apparatus 100 on a target block.

Motion information on the target block may be derived during inter prediction by each of the encoding apparatus 100 and the decoding apparatus 200 . Motion information may be derived using motion information of a reconstructed neighboring block, motion information of a collocated block, and/or motion information of a block adjacent to the collocated block.

For example, the encoding apparatus 100 or the decoding apparatus 200 performs prediction and/or motion compensation by using motion information of a spatial candidate and/or a temporal candidate as motion information of a target block. can be done A target block may mean a PU and/or a PU partition.

A spatial candidate may be a reconstructed block that is spatially adjacent to the target block.

A temporal candidate may be a reconstructed block corresponding to a target block in an already reconstructed collocated picture (col picture).

In inter prediction, the encoding apparatus 100 and the decoding apparatus 200 may improve encoding efficiency and decoding efficiency by using motion information of spatial candidates and/or temporal candidates. Motion information of spatial candidates may be referred to as spatial motion information. Motion information of temporal candidates may be referred to as temporal motion information.

Hereinafter, motion information of a spatial candidate may be motion information of a PU including the spatial candidate. Motion information of the temporal candidate may be motion information of a PU including the temporal candidate. Motion information of the candidate block may be motion information of a PU including the candidate block.

Inter prediction may be performed using a reference picture.

A reference picture may be at least one of a previous picture of the target picture or a subsequent picture of the target picture. A reference picture may refer to an image used for prediction of a target block.

In inter prediction, a region within a reference picture may be specified by using a reference picture index (or refIdx) indicating a reference picture and a motion vector to be described later. Here, a specified region within a reference picture may represent a reference block.

Inter prediction may select a reference picture and may select a reference block corresponding to a target block within the reference picture. In addition, inter prediction may generate a prediction block for a target block using the selected reference block.

Motion information may be derived during inter prediction by each of the encoding apparatus 100 and the decoding apparatus 200 .

A spatial candidate may be a block that 1) exists in the target picture, 2) has already been reconstructed through encoding and/or decoding, and 3) is located adjacent to or at a corner of the target block. Here, the block located at the corner of the target block may be a block vertically adjacent to a neighboring block horizontally adjacent to the target block or a block horizontally adjacent to a neighboring block vertically adjacent to the target block. "A block located at a corner of a target block" may have the same meaning as "a block adjacent to a corner of a target block". A "block located at a corner of a target block" may be included in a "block adjacent to the target block".

For example, spatial candidates include a reconstructed block located to the left of the target block, a reconstructed block located on top of the target block, a reconstructed block located at the lower left corner of the target block, and a reconstructed block located at the upper right corner of the target block. It can be a reconstructed block or a reconstructed block located in the upper left corner of the target block.

Each of the encoding apparatus 100 and the decoding apparatus 200 may identify a block existing at a position spatially corresponding to a target block within a col picture. A position of a target block in a target picture and a position of an identified block in a collocated picture may correspond to each other.

Each of the encoding apparatus 100 and the decoding apparatus 200 may determine a col block existing at a predefined relative position with respect to the identified block as a temporal candidate. The predefined relative position may be a position inside and/or outside the identified block.

For example, the collocated block may include a first collocated block and a second collocated block. When the coordinates of the identified block are (xP, yP) and the size of the identified block is (nPSW, nPSH), the first collocated block may be a block located at the coordinates (xP + nPSW, yP + nPSH). The second collocated block may be a block located at coordinates (xP + (nPSW >> 1), yP + (nPSH >> 1). The second call block may be selectively used when the first call block is unavailable.

The motion vector of the target block may be determined based on the motion vector of the collocated block. Each of the encoding apparatus 100 and the decoding apparatus 200 may scale a motion vector of a collocated block. A scaled motion vector of the collocated block may be used as a motion vector of the target block. Also, a motion vector of motion information of a temporal candidate stored in the list may be a scaled motion vector.

A ratio of the motion vector of the target block and the motion vector of the collocated block may be equal to the ratio of the first temporal distance to the second temporal distance. The first temporal distance may be the distance between the reference picture of the target block and the target picture. The second temporal distance may be a distance between a reference picture of a collocated block and a collocated picture.

A method of deriving motion information may change according to an inter prediction mode of a target block. For example, as inter prediction modes applied for inter prediction, an Advanced Motion Vector Predictor (AMVP) mode, a merge mode and a skip mode, a merge mode with motion vector difference, There may be sub-block merge mode, triangulation mode, inter-intra combined prediction mode, affine inter mode, and current picture reference mode. Merge mode may also be referred to as motion merge mode. Below, each of the modes is described in detail.

1) AMVP Mode

When the AMVP mode is used, the encoding apparatus 100 may search for similar blocks in the neighborhood of the target block. The encoding apparatus 100 may obtain a prediction block by performing prediction on a target block using motion information of a similar block found. The encoding apparatus 100 may encode a residual block that is a difference between a target block and a prediction block.

1-1) Creating a predicted motion vector candidate list

When the AMVP mode is used as the prediction mode, each of the encoding apparatus 100 and the decoding apparatus 200 may generate a predictive motion vector candidate list using a motion vector of a spatial candidate, a motion vector of a temporal candidate, and a zero vector. there is. The predicted motion vector candidate list may include one or more predicted motion vector candidates. At least one of the motion vector of the spatial candidate, the motion vector of the temporal candidate, and the zero vector may be determined and used as the predicted motion vector candidate.

Hereinafter, the terms “predicted motion vector (candidate)” and “motion vector (candidate)” may be used interchangeably and may be used interchangeably.

Hereinafter, the terms “predicted motion vector candidate” and “AMVP candidate” may be used interchangeably and may be used interchangeably.

Hereinafter, the terms “prediction motion vector candidate list” and “AMVP candidate list” may be used in the same meaning and may be used interchangeably.

Spatial candidates may include reconstructed spatial neighboring blocks. In other words, the motion vector of the reconstructed neighboring block may be referred to as a spatial prediction motion vector candidate.

Temporal candidates may include a collocated block and a block adjacent to the collocated block. In other words, a motion vector of a collocated block or a motion vector of a block adjacent to the collocated block may be referred to as a temporal prediction motion vector candidate.

The zero vector may be a (0, 0) motion vector.

The predictive motion vector candidate may be a motion vector predictor for motion vector prediction. Also, in the encoding apparatus 100, the predicted motion vector candidate may be an initial motion vector search position.

1-2) Motion vector search using a predictive motion vector candidate list

The encoding apparatus 100 may determine a motion vector to be used for encoding a target block within a search range by using the predictive motion vector candidate list. Also, the encoding apparatus 100 may determine a predictive motion vector candidate to be used as a predictive motion vector of the target block from among predictive motion vector candidates in the predictive motion vector candidate list.

A motion vector to be used for encoding a target block may be a motion vector that can be encoded with minimal cost.

Also, the encoding apparatus 100 may determine whether to use the AMVP mode in encoding the target block.

1-3) Transmission of inter prediction information

The encoding apparatus 100 may generate a bitstream including inter prediction information required for inter prediction. The decoding apparatus 200 may perform inter prediction on a target block using inter prediction information of a bitstream.

The inter-prediction information includes 1) mode information indicating whether the AMVP mode is used, 2) predictive motion vector index, 3) motion vector difference (MVD), 4) reference direction, and 5) reference picture index. can do.

Hereinafter, the terms “predicted motion vector index” and “AMVP index” may be used interchangeably and may be used interchangeably.

Also, the inter prediction information may include a residual signal.

When the mode information indicates that the AMVP mode is used, the decoding apparatus 200 may obtain a predicted motion vector index, a motion vector difference, a reference direction, and a reference picture index from a bitstream through entropy decoding.

The predictive motion vector index may indicate a predictive motion vector candidate used for prediction of a target block among predictive motion vector candidates included in the predictive motion vector candidate list.

1-4) Inter prediction of AMVP mode using inter prediction information

The decoding apparatus 200 may derive a predicted motion vector candidate using the predicted motion vector candidate list, and may determine motion information of a target block based on the derived motion vector predicted candidate.

The decoding apparatus 200 may determine a motion vector candidate for a target block from among motion vector predictor candidates included in a predictor motion vector candidate list by using the predictor motion vector index. The decoding apparatus 200 may select a predicted motion vector candidate indicated by a predicted motion vector index as a predicted motion vector of a target block from among motion vector predicted candidates included in the predicted motion vector candidate list.

The encoding apparatus 100 may generate an entropy-encoded predicted motion vector index by applying entropy encoding to the predicted motion vector index, and may generate a bitstream including the entropy-encoded predicted motion vector index. The entropy-encoded predicted motion vector index may be signaled from the encoding apparatus 100 to the decoding apparatus 200 through a bitstream. The decoding apparatus 200 may extract an entropy-encoded predictive motion vector index from a bitstream and obtain the predictive motion vector index by applying entropy decoding to the entropy-encoded predictive motion vector index.

A motion vector actually used for inter prediction of a target block may not coincide with a predicted motion vector. MVD may be used to represent a difference between a motion vector actually used for inter prediction of a target block and a predicted motion vector. The encoding apparatus 100 may derive a predicted motion vector similar to a motion vector actually used for inter prediction of a target block in order to use an MVD having a size as small as possible.

MVD may be a difference between a motion vector of a target block and a predicted motion vector. The encoding apparatus 100 may calculate the MVD and generate an entropy-encoded MVD by applying entropy encoding to the MVD. The encoding apparatus 100 may generate a bitstream including an entropy-encoded MDV.

The MVD may be transmitted from the encoding device 100 to the decoding device 200 through a bitstream. The decoding apparatus 200 may extract the entropy-encoded MVD from the bitstream and obtain the MVD by applying entropy decoding to the entropy-encoded MVD.

The decoding apparatus 200 may derive the motion vector of the target block by adding the MVD and the predicted motion vector. In other words, the motion vector of the target block derived by the decoding apparatus 200 may be the sum of the MVD and the motion vector candidate.

Also, the encoding apparatus 100 may generate entropy-encoded MVD resolution information by applying entropy encoding to the calculated MVD resolution information, and may generate a bitstream including the entropy-encoded MVD resolution information. The decoding apparatus 200 may extract entropy-encoded MVD resolution information from a bitstream, and obtain MVD resolution information by applying entropy decoding to the entropy-encoded MVD resolution information. The decoding apparatus 200 may adjust the resolution of the MVD using the MVD resolution information.

Meanwhile, the encoding device 100 may calculate the MVD based on the affine model. The decoding apparatus 200 may derive an affine control motion vector of the target block through the sum of the MVD and the affine control motion vector candidate, and may derive a motion vector for a subblock using the affine control motion vector. there is.

A reference direction may indicate a reference picture list used for prediction of a target block. For example, the reference direction may indicate one of a reference picture list L0 and a reference picture list L1.

The reference direction indicates only a reference picture list used for prediction of the target block, and may not indicate that directions of reference pictures are limited to a forward direction or a backward direction. In other words, each of the reference picture list L0 and the reference picture list L1 may include forward and/or backward pictures.

That the reference direction is uni-direction may mean that one reference picture list is used. Bi-direction of the reference direction may mean that two reference picture lists are used. In other words, the reference direction can indicate one of: that only the reference picture list L0 is used, that only the reference picture list L1 is used, and two reference picture lists.

The reference picture index may indicate a reference picture used for prediction of a target block among reference pictures of a reference picture list. The encoding apparatus 100 may generate an entropy-coded reference picture index by applying entropy encoding to the reference picture index, and may generate a bitstream including the entropy-coded reference picture index. The entropy-encoded reference picture index may be signaled from the encoding device 100 to the decoding device 200 through a bitstream. The decoding apparatus 200 may extract an entropy-encoded reference picture index from a bitstream and obtain the reference picture index by applying entropy decoding to the entropy-encoded reference picture index.

When two reference picture lists are used for prediction of a target block. One reference picture index and one motion vector may be used for each reference picture list. Also, when two reference picture lists are used for prediction of a target block, two prediction blocks may be specified for the target block. For example, a (final) prediction block of the target block may be generated through an average or a weighted-sum of two prediction blocks of the target block.

The motion vector of the target block may be derived by the predicted motion vector index, MVD, reference direction, and reference picture index.

The decoding apparatus 200 may generate a prediction block for the target block based on the derived motion vector and the reference picture index. For example, the prediction block may be a reference block indicated by a derived motion vector in a reference picture indicated by a reference picture index.

By encoding the predicted motion vector index and the MVD without encoding the motion vector of the target block, the amount of bits transmitted from the encoding apparatus 100 to the decoding apparatus 200 can be reduced and encoding efficiency can be improved.

Motion information of a neighboring block reconstructed with respect to the target block may be used. In a specific inter-prediction mode, the encoding apparatus 100 may not separately encode the motion information of the target block itself. Motion information of the target block is not encoded, and other information capable of inducing motion information of the target block through motion information of the reconstructed neighboring block may be encoded instead. As other information is encoded instead, the amount of bits transmitted to the decoding apparatus 200 can be reduced and encoding efficiency can be improved.

For example, as an inter prediction mode in which motion information of the target block is not directly encoded, there may be a skip mode and/or a merge mode. In this case, the encoding apparatus 100 and the decoding apparatus 200 may use an identifier and/or an index indicating which unit's motion information among reconstructed neighboring units is used as the motion information of the target unit.

2) Merge Mode

As a method of deriving motion information of a target block, there is a merge. Merge may refer to merging of motions of a plurality of blocks. Merge may mean applying motion information of one block to another block as well. In other words, merge mode may refer to a mode in which motion information of a target block is derived from motion information of neighboring blocks.

When merge mode is used, the encoding apparatus 100 may perform prediction of motion information of a target block by using motion information of a spatial candidate and/or motion information of a temporal candidate. The spatial candidate may include a reconstructed spatial neighboring block spatially adjacent to the target block. A spatial neighboring block may include a left neighboring block and an upper neighboring block. Temporal candidates may include call blocks. The terms "spatial candidate" and "spatial merge candidate" may be used interchangeably and may be used interchangeably. The terms "temporal candidate" and "temporal merge candidate" may be used interchangeably and may be used interchangeably.

The encoding apparatus 100 may obtain a prediction block through prediction. The encoding apparatus 100 may encode a residual block that is a difference between a target block and a prediction block.

2-1) Creating a merge candidate list

When the merge mode is used, each of the encoding apparatus 100 and the decoding apparatus 200 may generate a merge candidate list using motion information of spatial candidates and/or motion information of temporal candidates. Motion information may include 1) a motion vector, 2) a reference picture index, and 3) a reference direction. The reference direction can be unidirectional or bidirectional. The reference direction may mean an inter prediction indicator.

The merge candidate list may include merge candidates. A merge candidate may be motion information. In other words, the merge candidate list may be a list in which motion information is stored.

Merge candidates may be motion information such as temporal candidates and/or spatial candidates. In other words, the merge candidate list may include motion information such as a temporal candidate and/or a spatial candidate.

Also, the merge candidate list may include a new merge candidate generated by a combination of merge candidates already existing in the merge candidate list. In other words, the merge candidate list may include new motion information generated by combining motion information already existing in the merge candidate list.

Also, the merge candidate list may include a history-based merge candidate. A history-based merge candidate may be motion information of a block encoded and/or decoded prior to a target block.

Also, the merge candidate list may include a merge candidate based on an average of two merge candidates.

Merge candidates may be specified modes for deriving inter prediction information. A merge candidate may be information indicating a specific mode for deriving inter prediction information. Inter prediction information of a target block may be derived according to a specified mode indicated by a merge candidate. In this case, the specified mode may include a process of deriving a series of inter prediction information. This specified mode may be an inter prediction information derivation mode or a motion information derivation mode.

Inter prediction information of a target block may be derived according to a mode indicated by a merge candidate selected by a merge index among merge candidates in the merge candidate list.

For example, the motion information derivation modes in the merge candidate list may be at least one of 1) a motion information derivation mode in units of sub-blocks and 2) an affine motion information derivation mode.

Also, the merge candidate list may include motion information of a zero vector. A zero vector may be referred to as a zero merge candidate.

In other words, the motion information in the merge candidate list is: 1) motion information of a spatial candidate, 2) motion information of a temporal candidate, 3) motion information generated by a combination of motion information already existing in the merge candidate list, 4) zero vector may be at least one of

Motion information may include 1) a motion vector, 2) a reference picture index, and 3) a reference direction. A reference direction may be referred to as an inter prediction indicator. The reference direction can be unidirectional or bidirectional. A unidirectional reference direction may represent L0 prediction or L1 prediction.

The merge candidate list may be generated before prediction by merge mode is performed.

The number of merge candidates in the merge candidate list may be predefined. The encoding device 100 and the decoding device 200 may add merge candidates to the merge candidate list according to a predefined method and a predefined order so that the merge candidate list includes a predefined number of merge candidates. The merge candidate list of the encoding device 100 and the merge candidate list of the decoding device 200 may be the same through a predefined method and a predefined ranking.

Merge may be applied in units of CUs or units of PUs. When merging is performed in units of CUs or PUs, the encoding apparatus 100 may transmit a bitstream including predefined information to the decoding apparatus 200. For example, the predefined information includes 1) information indicating whether merging is to be performed for each block partition, 2) which block to merge with among blocks that are spatial candidates and/or temporal candidates for the target block. It may contain information on whether

2-2) Motion vector search using merge candidate list

The encoding apparatus 100 may determine a merge candidate to be used for encoding a target block. For example, the encoding apparatus 100 may perform predictions on a target block using merge candidates of the merge candidate list and generate residual blocks for the merge candidates. The encoding apparatus 100 may use a merge candidate that requires the least cost in encoding the prediction and residual blocks for encoding the target block.

Also, the encoding apparatus 100 may determine whether to use merge mode in encoding a target block.

2-3) Transmission of inter prediction information

The encoding apparatus 100 may generate a bitstream including inter prediction information required for inter prediction. The encoding apparatus 100 may generate entropy-encoded inter-prediction information by performing entropy encoding on the inter-prediction information, and may transmit a bitstream including the entropy-encoded inter-prediction information to the decoding apparatus 200. Entropy-encoded inter prediction information may be signaled from the encoding apparatus 100 to the decoding apparatus 200 through a bitstream. The decoding apparatus 200 may extract entropy-encoded inter prediction information from a bitstream and obtain inter prediction information by performing entropy decoding on the entropy-encoded inter prediction information.

The decoding apparatus 200 may perform inter prediction on a target block using inter prediction information of a bitstream.

The inter prediction information may include 1) mode information indicating whether merge mode is used, 2) merge index, and 3) correction information.

Also, the inter prediction information may include a residual signal.

The decoding apparatus 200 may obtain the merge index from the bitstream only when the mode information indicates that the merge mode is used.

Mode information may be a merge flag. A unit of mode information may be a block. Information about a block may include mode information, and the mode information may indicate whether merge mode is applied to the block.

The merge index may indicate a merge candidate used for prediction of a target block among merge candidates included in the merge candidate list. Alternatively, the merge index may indicate which block among neighboring blocks that are spatially or temporally adjacent to the target block is merged with.

The encoding apparatus 100 may select a merge candidate having the highest encoding performance among merge candidates included in the merge candidate list, and set a merge index value to indicate the selected merge candidate.

The correction information may be information used for motion vector correction. The encoding device 100 may generate correction information. The decoding apparatus 200 may correct the motion vector of the merge candidate selected by the merge index based on the correction information.

The correction information may include at least one of information indicating whether correction is performed, correction direction information, and correction size information. A prediction mode for correcting a motion vector based on signaled correction information may be referred to as a merge mode with motion vector difference.

2-4) Inter prediction of merge mode using inter prediction information

The decoding apparatus 200 may perform prediction on a target block by using a merge candidate indicated by a merge index among merge candidates included in the merge candidate list.

The motion vector of the target block may be specified by the motion vector of the merge candidate indicated by the merge index, the reference picture index, and the reference direction.

3) Skip Mode

The skip mode may be a mode in which motion information of a spatial candidate or motion information of a temporal candidate is applied to a target block as it is. Also, the skip mode may be a mode not using a residual signal. That is to say, when skip mode is used, the reconstructed block may be the same as the predicted block.

A difference between merge mode and skip mode may be transmission or use of a residual signal. In other words, skip mode may be similar to merge mode except that no residual signal is transmitted or used.

When the skip mode is used, the encoding apparatus 100 transmits information indicating which block's motion information among spatial or temporal candidate blocks is used as the motion information of the target block to the decoding apparatus 200 through a bitstream. can transmit The encoding apparatus 100 may generate entropy-encoded information by performing entropy encoding on such information, and may signal the entropy-encoded information to the decoding apparatus 200 through a bitstream. The decoding apparatus 200 may extract entropy-encoded information from a bitstream and obtain information by performing entropy decoding on the entropy-encoded information.

Also, when the skip mode is used, the encoding device 100 may not transmit other syntax element information such as MVD to the decoding device 200. For example, when the skip mode is used, the encoding apparatus 100 may not signal a syntax element related to at least one of the MVD, the coded block flag, and the transform coefficient level to the decoding apparatus 200.

3-1) Creation of merge candidate list

Skip mode can also use a merge candidate list. In other words, the merge candidate list can be used in both merge mode and skip mode. In this regard, the merge candidate list may be referred to as a "skip candidate list" or a "merge/skip candidate list".

Alternatively, skip mode may use a separate candidate list different from merge mode. In this case, in the description below, the merge candidate list and the merge candidate may be replaced with the skip candidate list and the skip candidate, respectively.

The merge candidate list may be generated before prediction by skip mode is performed.

3-2) Motion vector search using merge candidate list

The encoding apparatus 100 may determine a merge candidate to be used for encoding a target block. For example, the encoding apparatus 100 may perform predictions on a target block using merge candidates of a merge candidate list. The encoding apparatus 100 may use a merge candidate requiring a minimum prediction cost for encoding a target block.

Also, the encoding apparatus 100 may determine whether to use a skip mode in encoding a target block.

3-3) Transmission of inter prediction information

The encoding apparatus 100 may generate a bitstream including inter prediction information required for inter prediction. The decoding apparatus 200 may perform inter prediction on a target block using inter prediction information of a bitstream.

The inter-prediction information may include 1) mode information indicating whether a skip mode is used and 2) a skip index.

The skip index may be the same as the aforementioned merge index.

When the skip mode is used, the target block may be coded without a residual signal. Inter prediction information may not include a residual signal. Alternatively, the bitstream may not include a residual signal.

The decoding apparatus 200 may obtain the skip index from the bitstream only when the mode information indicates that the skip mode is used. As mentioned above, the merge index and skip index may be the same. The decoding apparatus 200 may obtain the skip index from the bitstream only when the mode information indicates that merge mode or skip mode is used.

The skip index may indicate a merge candidate used for prediction of a target block among merge candidates included in the merge candidate list.

3-4) Skip Mode Inter Prediction Using Inter Prediction Information

The decoding apparatus 200 may perform prediction on a target block by using a merge candidate indicated by a skip index among merge candidates included in the merge candidate list.

A motion vector of a target block may be specified by a motion vector of a merge candidate indicated by a skip index, a reference picture index, and a reference direction.

4) Current picture reference mode

The current picture reference mode may mean a prediction mode using a pre-reconstructed region in a target picture to which a target block belongs.

A motion vector may be used to specify the pre-reconstructed area. Whether the target block is encoded in the current picture reference mode can be determined using the reference picture index of the target block.

A flag or index indicating whether the target block is a block encoded in the current picture reference mode may be signaled from the encoding device 100 to the decoding device 200. Alternatively, whether the target block is a block coded in the current picture reference mode may be inferred through a reference picture index of the target block.

When the target block is coded in the current picture reference mode, the target picture may exist at a fixed position or an arbitrary position in the reference picture list for the target block.

For example, the fixed position may be a position where the value of the reference picture index is 0 or the last position.

If the target picture exists at an arbitrary position in the reference picture list, a separate reference picture index indicating such an arbitrary position may be signaled from the encoding apparatus 100 to the decoding apparatus 200.

5) Subblock merge mode

A sub-block merge mode may mean a mode for deriving motion information for a sub-block of a CU.

When the sub-block merge mode is applied, motion information of a call sub-block of a target sub-block in a reference image (ie, a sub-block based temporal merge candidate) and/or an affine control point motion vector A subblock merge candidate list may be generated using an affine control point motion vector merge candidate.

6) Triangle partition mode

In the triangular division mode, divided target blocks may be generated by dividing the target block in a diagonal direction. For each divided target block, motion information of each divided target block may be derived, and prediction samples for each divided target block may be derived using the derived motion information. A prediction sample of the target block may be derived through a weighted sum of prediction samples of the divided target blocks.

7) Inter-intra combined prediction mode

The inter-intra combined prediction mode may be a mode in which a prediction sample of a target block is derived by using a weighted sum of prediction samples generated by inter prediction and prediction samples generated by intra prediction.

In the aforementioned modes, the decoding apparatus 200 may perform self-correction on the derived motion information. For example, the decoding apparatus 200 may search a specific area based on a reference block indicated by the derived motion information to search for motion information having a minimum sum of absolute differences (SAD). and the searched motion information may be derived as corrected motion information.

In the aforementioned modes, the decoding apparatus 200 may perform compensation for prediction samples derived through inter prediction using an optical flow.

In the aforementioned AMVP mode, merge mode, and skip mode, motion information to be used for prediction of a target block among motion information in the list may be specified through an index of the list.

In order to improve encoding efficiency, the encoding apparatus 100 may signal only the index of an element that causes the least cost in inter prediction of a target block among elements in the list. The encoding device 100 may encode the index and signal the encoded index.

Accordingly, the aforementioned lists (ie, the predicted motion vector candidate list and the merge candidate list) may have to be derived in the same manner based on the same data in the encoding apparatus 100 and the decoding apparatus 200. Here, the same data may include a reconstructed picture and a reconstructed block. Also, in order to specify an element by index, the order of the elements within the list may need to be constant.

10 shows spatial candidates according to an example.

In Fig. 10, the positions of spatial candidates are shown.

A large block in the middle may represent a target block. Five small blocks may represent spatial candidates.

Coordinates of the target block may be (xP, yP), and the size of the target block may be (nPSW, nPSH).

The spatial candidate A ₀ may be a block adjacent to the lower left corner of the target block. A ₀ may be a block occupying pixels of coordinates (xP - 1, yP + nPSH).

Spatial candidate A ₁ may be a block adjacent to the left of the target block. A ₁ may be the lowest block among blocks adjacent to the left side of the target block. Alternatively, A ₁ may be a block adjacent to the top of A ₀ . A ₁ may be a block occupying pixels of coordinates (xP - 1, yP + nPSH - 1).

Spatial candidate B ₀ may be a block adjacent to the upper right corner of the target block. B ₀ may be a block occupying pixels of coordinates (xP + nPSW, yP - 1).

Spatial candidate B ₁ may be a block adjacent to the top of the target block. B ₁ may be the rightmost block among blocks adjacent to the top of the target block. Alternatively, B ₁ may be a block adjacent to the left side of B ₀ . B ₁ may be a block occupying pixels of coordinates (xP + nPSW - 1, yP - 1).

Spatial candidate B ₂ may be a block adjacent to the upper left corner of the target block. B ₂ may be a block occupying pixels of coordinates (xP - 1, yP - 1).

Determination of the availability of spatial and temporal candidates

In order to include the motion information of the spatial candidate or the motion information of the temporal candidate in the list, it must be determined whether the motion information of the spatial candidate or the motion information of the temporal candidate is available.

Hereinafter, a candidate block may include a spatial candidate and a temporal candidate.

For example, the above determination may be made by sequentially applying steps 1) to 4) below.

Step 1) If the PU including the candidate block is outside the boundary of the picture, the availability of the candidate block may be set to false. "Availability is set to false" may mean the same as "availability is set to unavailability".

Step 2) If the PU including the candidate block is outside the boundary of the slice, the availability of the candidate block may be set to false. If the target block and the candidate block are located in different slices, the availability of the candidate block may be set to false.

Step 3) If the PU including the candidate block is outside the boundary of the tile, the availability of the candidate block may be set to false. If the target block and the candidate block are located in different tiles, the availability of the candidate block may be set to false.

Step 4) If the prediction mode of the PU including the candidate block is an intra prediction mode, availability of the candidate block may be set to false. If the PU containing the candidate block does not use inter prediction, the availability of the candidate block may be set to false.

11 illustrates an order of adding motion information of spatial candidates to a merge list according to an example.

As shown in FIG. 11 , in adding motion information of spatial candidates to the merge list, the order of A ₁ , B ₁ , B ₀ , A ₀ and B ₂ may be used. That is, in the order of A ₁ , B ₁ , B ₀ , A ₀ and B ₂ , motion information of available spatial candidates may be added to the merge list.

Merge list derivation method in merge mode and skip mode

As described above, the maximum number of merge candidates in the merge list may be set. The set maximum number is indicated by N. The set number may be transmitted from the encoding device 100 to the decoding device 200. A slice header of a slice may include N. In other words, the maximum number of merge candidates of the merge list for the target block of the slice may be set by the slice header. For example, the value of N may be 5 by default.

Motion information (ie, merge candidates) may be added to the merge list in the order of steps 1) to 4) below.

Step 1) Available spatial candidates among spatial candidates may be added to the merge list. Motion information of available spatial candidates may be added to the merge list in the order shown in FIG. 11 . In this case, if motion information of an available spatial candidate overlaps with other motion information already existing in the merge list, the motion information may not be added to the merge list. Checking whether it overlaps with other motion information existing in the list can be abbreviated as "redundancy check".

A maximum of N pieces of motion information may be added.

Step 2) If the number of pieces of motion information in the merge list is smaller than N and a temporal candidate is available, the motion information of the temporal candidate may be added to the merge list. In this case, if motion information of an available temporal candidate overlaps with other motion information already existing in the merge list, the motion information may not be added to the merge list.

Step 3) If the number of pieces of motion information in the merge list is smaller than N and the type of the target slice is "B", the combined motion information generated by combined bi-prediction is added to the merge list. can

The target slice may be a slice including the target block.

The combined motion information may be a combination of L0 motion information and L1 motion information. The L0 motion information may be motion information referring only to the reference picture list L0. The L1 motion information may be motion information referring only to the reference picture list L1.

Within the merge list, there may be one or more L0 motion information. Also, within the merge list, there may be one or more L1 motion information.

The combined motion information may be one or more. In generating the combined motion information, which L0 motion information and which L1 motion information among one or more pieces of L0 motion information and one or more pieces of L1 motion information are to be used may be predefined. One or more pieces of combined motion information may be generated in a predefined order by combined bidirectional prediction using pairs of different pieces of motion information in a merge list. One of the pairs of different motion information may be L0 motion information and the other may be L1 motion information.

For example, the combined motion information added with the highest priority may be a combination of L0 motion information having a merge index of 0 and L1 motion information having a merge index of 1. If the motion information with a merge index of 0 is not L0 motion information or the motion information with a merge index of 1 is not L1 motion information, the combined motion information may not be generated and added. Motion information added next may be a combination of L0 motion information having a merge index of 1 and L1 motion information having a merge index of 0. The following specific combinations may follow other combinations in the field of encoding/decoding of video.

In this case, if the combined motion information overlaps with other motion information already existing in the merge list, the combined motion information may not be added to the merge list.

Step 4) If the number of pieces of motion information in the merge list is smaller than N, zero vector motion information may be added to the merge list.

The zero vector motion information may be motion information in which a motion vector is a zero vector.

One or more zero vector motion information may be provided. Reference picture indexes of one or more zero vector motion information may be different from each other. For example, the value of the reference picture index of the first zero vector motion information may be 0. The value of the reference picture index of the second zero vector motion information may be 1.

The number of zero vector motion information may be equal to the number of reference pictures in the reference picture list.

The reference direction of the zero vector motion information may be bidirectional. Both motion vectors may be zero vectors. The number of zero vector motion information may be the smaller of the number of reference pictures in the reference picture list L0 and the number of reference pictures in the reference picture list L1. Alternatively, when the number of reference pictures in the reference picture list L0 and the number of reference pictures in the reference picture list L1 are different from each other, a unidirectional reference direction may be used for a reference picture index applicable to only one reference picture list.

The encoding apparatus 100 and/or the decoding apparatus 200 may sequentially add zero vector motion information to the merge list while changing the reference picture index.

If the zero vector motion information overlaps with other motion information already present in the merge list, the zero vector motion information may not be added to the merge list.

The order of steps 1) to 4) described above is merely exemplary, and the order between steps may be interchanged. Also, some of the steps may be omitted according to predefined conditions.

Derivation method of predictive motion vector candidate list in AMVP mode

The maximum number of motion vector predictor candidates in the predictor motion vector candidate list may be predefined. The predefined maximum number is denoted by N. For example, the predefined maximum number may be 2.

Motion information (ie, predicted motion vector candidates) may be added to the predicted motion vector candidate list in the order of steps 1) to 3) below.

Step 1) Among the spatial candidates, available spatial candidates may be added to the predicted motion vector candidate list. Spatial candidates may include a first spatial candidate and a second spatial candidate.

The first spatial candidate may be one of A ₀ , A ₁ , scaled A ₀ and scaled A ₁ . The second spatial candidate may be one of B ₀ , B ₁ , B ₂ , scaled B ₀ , scaled B ₁ , and scaled B ₂ .

Motion information of available spatial candidates may be added to the predicted motion vector candidate list in the order of the first spatial candidate and the second spatial candidate. In this case, when motion information of an available spatial candidate overlaps with other motion information already existing in the motion vector predictor candidate list, the motion information may not be added to the predictor motion vector candidate list. In other words, when the value of N is 2, if the motion information of the second spatial candidate is identical to the motion information of the first spatial candidate, the motion information of the second spatial candidate may not be added to the predicted motion vector candidate list.

A maximum of N pieces of motion information may be added.

Step 2) If the number of pieces of motion information in the predicted motion vector candidate list is smaller than N and a temporal candidate is available, the motion information of the temporal candidate may be added to the predicted motion vector candidate list. In this case, when motion information of an available temporal candidate overlaps with other motion information already existing in the motion vector predictor candidate list, the motion information may not be added to the predictor motion vector candidate list.

Step 3) If the number of pieces of motion information in the predicted motion vector candidate list is smaller than N, zero vector motion information may be added to the predicted motion vector candidate list.

One or more zero vector motion information may be provided. Reference picture indexes of one or more zero vector motion information may be different from each other.

The encoding apparatus 100 and/or the decoding apparatus 200 may sequentially add zero vector motion information to the predictive motion vector candidate list while changing the reference picture index.

When the zero vector motion information overlaps with other motion information already existing in the predicted motion vector candidate list, the zero vector motion information may not be added to the predicted motion vector candidate list.

The description of the zero vector motion information described above for the merge list can also be applied to the zero vector motion information. Redundant descriptions are omitted.

The order of steps 1) to 3) described above is merely exemplary, and the order between steps may be interchanged. Also, some of the steps may be omitted according to predefined conditions.

12 illustrates a process of transformation and quantization according to an example.

As shown in FIG. 12, a quantized level may be generated by performing a transform and/or quantization process on the residual signal.

A residual signal may be generated as a difference between an original block and a prediction block. Here, the prediction block may be a block generated by intra prediction or inter prediction.

The residual signal may be converted into the frequency domain through a transform process that is part of the quantization process.

Transformation kernels used for transformation may include various DCT kernels such as Discrete Cosine Transform (DCT) type 2 (DCT-II) and Discrete Sine Transform (DST) kernels. .

These transform kernels may perform a separable transform or a 2D (2D) non-separable transform on the residual signal. The separable transform may be a transform that performs a one-dimensional (1D) transform on the residual signal in each of a horizontal direction and a vertical direction.

DCT type and DST type adaptively used for 1D conversion may include DCT-V, DCT-VIII, DST-I and DST-VII in addition to DCT-II as shown in Table 3 and Table 4 below, respectively. there is.

[Table 3]

[Table 4]

As shown in Tables 3 and 4, a transform set may be used in deriving the DCT type or DST type to be used for transformation. Each transform set may include a plurality of transform candidates. Each transformation candidate may be a DCT type or a DST type.

Table 5 below shows an example of a transform set applied in the horizontal direction and a transform set applied in the vertical direction according to the intra prediction mode.

[Table 5]

In Table 5, vertical transform set numbers and horizontal transform set numbers applied to the horizontal direction of the residual signal are indicated according to the intra prediction mode of the target block.

As illustrated in Table 5, transform sets applied in the horizontal and vertical directions may be predefined according to the intra prediction mode of the target block. The encoding apparatus 100 may perform transform and inverse transform on the residual signal using a transform included in a transform set corresponding to the intra prediction mode of the target block. Also, the decoding apparatus 200 may perform an inverse transform on the residual signal using a transform included in a transform set corresponding to the intra prediction mode of the target block.

In these transforms and inverse transforms, the set of transforms applied to the residual signal may be determined as exemplified in Tables 3, 4, and 5, and may not be signaled. Transformation indication information may be signaled from the encoding device 100 to the decoding device 200. The transform indication information may be information indicating which transform candidate among a plurality of transform candidates included in a transform set applied to the residual signal is used.

For example, when the size of the target block is 64x64 or less, transform sets each having three transforms may be configured according to the intra prediction mode. An optimal transform method can be selected among all 9 multiple transform methods resulting from a combination of three transforms in the horizontal direction and three transforms in the vertical direction. Encoding efficiency can be improved by encoding and/or decoding the residual signal using such an optimal conversion method.

In this case, information on which transform among transforms belonging to the transform set is used for at least one of the vertical transform and the horizontal transform may be entropy encoded and/or decoded. A truncated unary binarization may be used to encode and/or decode this information.

As described above, a method using various transforms may be applied to a residual signal generated by intra prediction or inter prediction.

Transformation may include at least one of a primary transformation and a secondary transformation. A transform coefficient may be generated by performing a primary transform on the residual signal, and a secondary transform coefficient may be generated by performing a secondary transform on the transform coefficient.

A primary transformation may be named primary. Also, the primary transform may be referred to as an adaptive multiple transform (AMT). As described above, AMT may mean that different transforms are applied to each of the 1D directions (ie, the vertical direction and the horizontal direction).

The secondary transform may be a transform for improving the energy concentration of the transform coefficient generated by the primary transform. Like the first-order transformation, the second-order transformation may be a separable transformation or a non-separable transformation. The non-separable transform may be a non-separable secondary transform (NSST).

The primary transformation may be performed using at least one of a plurality of predefined transformation methods. For example, a plurality of predefined transform methods include a Discrete Cosine Transform (DCT), a Discrete Sine Transform (DST), and a Karhunen-Loeve Transform (KLT) based transform. can include

Also, the primary transform may be a transform having various transform types according to a kernel function defining DCT or DST.

For example, the transform type includes: 1) prediction mode of the target block (eg, one of intra prediction and inter prediction), 2) size of the target block, 3) shape of the target block, 4) intra prediction mode of the target block , 5) components of the target block (eg, one of a luma component and a chroma component) and 6) a partition type applied to the target block (eg, Quad Tree (QT), Binary Tree (BT)). ) and one of a ternary tree (TT)).

For example, the primary transform includes transforms such as DCT-2, DCT-5, DCT-7, DST-7, DST-1, DST-8 and DCT-8 according to the transform kernel presented in Table 6 below. can do. In Table 6, various transform types and transform kernel functions for Multiple Transform Selection (MTS) are illustrated.

MTS may mean that a combination of one or more DCT and/or DST transform kernels is selected for horizontal and/or vertical transform of the residual signal.

[Table 6]

In Table 6, i and j may be integer values greater than or equal to 0 and less than or equal to N-1.

A secondary transform may be performed on transform coefficients generated by performing the primary transform.

As in the first-order transform, a set of transforms can be defined in the second-order transform. Methods for deriving and/or determining a set of transforms, such as those described above, may be applied to first-order as well as second-order transforms.

A primary transform and a secondary transform can be determined for a specified object.

For example, the first transform and the second transform may be applied to one or more signal components of a luma component and a chroma component. Whether to apply the primary transform and/or the secondary transform may be determined according to at least one of coding parameters for a target block and/or a neighboring block. For example, whether to apply the first transform and/or the second transform may be determined by the size and/or shape of the target block.

In the encoding device 100 and the decoding device 200, transformation information indicating a transformation method used for a target may be derived by using specified information.

For example, the transformation information may include an index of a transformation to be used for primary transformation and/or secondary transformation. Alternatively, the transform information may indicate that the primary transform and/or the secondary transform are not used.

For example, when the target of the primary transform and the secondary transform is a target block, the transform method(s) applied to the primary transform and/or the secondary transform indicated by transform information may be applied to the target block and/or neighboring blocks. It may be determined according to at least one of the coding parameters for

Alternatively, transformation information indicating a transformation method for a specified target may be signaled from the encoding apparatus 100 to the decoding apparatus 200.

For example, whether or not a primary transform is used for one CU, an index indicating the primary transform, whether or not a secondary transform is used, and an index indicating the secondary transform may be derived as transformation information in the decoding apparatus 200 there is. Alternatively, conversion information indicating whether or not the primary transform is used for one CU, an index indicating the primary transform, whether or not a secondary transform is used, and an index indicating the secondary transform may be signaled.

A quantized transform coefficient (ie, a quantized level) may be generated by performing quantization on a residual signal or a result generated by performing the primary transform and/or the secondary transform.

13 illustrates diagonal scanning according to an example.

14 illustrates horizontal scanning according to an example.

15 illustrates vertical scanning according to an example.

The quantized transform coefficients may be scanned according to at least one of (up-right) diagonal scanning, vertical scanning, and horizontal scanning according to at least one of an intra prediction mode, a block size, and a block shape. A block may be a transform unit.

Each scanning can start at a specified start point and end at a specified end point.

For example, quantized transform coefficients may be changed into a one-dimensional vector form by scanning coefficients of a block using the diagonal scanning of FIG. 13 . Alternatively, the horizontal scanning of FIG. 14 or the vertical scanning of FIG. 15 may be used instead of diagonal scanning according to the block size and/or intra prediction mode.

Vertical scanning may be scanning two-dimensional block form coefficients in a column direction. Horizontal scanning may be scanning two-dimensional block form coefficients in a row direction.

In other words, which of diagonal scanning, vertical scanning, and horizontal scanning is to be used may be determined according to the size of the block and/or the inter-prediction mode.

As shown in FIGS. 13, 14, and 15, quantized transform coefficients may be scanned along a diagonal, horizontal, or vertical direction.

Quantized transform coefficients may be expressed in block form. A block may include a plurality of sub-blocks. Each sub-block may be defined according to a minimum block size or a minimum block shape.

In scanning, a scanning order according to a type or direction of scanning may be applied to subblocks first. Also, a scanning order according to a scanning direction may be applied to quantized transform coefficients in a sub-block.

For example, as shown in FIGS. 13, 14, and 15, when the size of the target block is 8x8, transform coefficients quantized by primary transform, secondary transform, and quantization of the residual signal of the target block are can be created Thereafter, one of three scanning orders may be applied to the four 4x4 subblocks, and quantized transform coefficients may be scanned according to the scanning order for each 4x4 subblock.

The encoding apparatus 100 may generate an entropy-encoded quantized transform coefficient by performing entropy encoding on the scanned quantized transform coefficients, and may generate a bitstream including the entropy-coded quantized transform coefficients. .

The decoding apparatus 200 may extract entropy-encoded quantized transform coefficients from a bitstream and generate quantized transform coefficients by performing entropy decoding on the entropy-coded quantized transform coefficients. Quantized transform coefficients may be arranged in a 2D block form through inverse scanning. At this time, as a reverse scanning method, at least one of (upper right) diagonal scan, vertical scan, and horizontal scan may be performed.

In the decoding apparatus 200, inverse quantization may be performed on quantized transform coefficients. Depending on whether the second-order inverse transform is performed, the second-order inverse transform may be performed on the result generated by performing the inverse quantization. Also, depending on whether the first-order inverse transform is performed, the first-order inverse transform may be performed on the result generated by performing the second-order inverse transform. A reconstructed residual signal may be generated by performing a first-order inverse transform on a result generated by performing a second-order inverse transform.

For a luma component reconstructed through intra prediction or inter prediction, inverse mapping of a dynamic range may be performed before in-loop filtering.

The dynamic range may be divided into 16 equal pieces, and a mapping function for each piece may be signaled. The mapping function may be signaled at the slice level or tile group level.

An inverse mapping function for performing inverse mapping may be derived based on the mapping function.

In-loop filtering, reference picture storage, and motion compensation may be performed in the inversely mapped region.

A prediction block generated through inter prediction may be converted into a mapped region by mapping using a mapping function, and the converted prediction block may be used to generate a reconstructed block. However, since intra prediction is performed in a mapped region, a prediction block generated by intra prediction can be used to generate a reconstructed block without mapping and/or inverse mapping.

For example, when the target block is a residual block of a chroma component, the residual block may be converted into an inversely mapped region by scaling the chroma component of the mapped region.

Whether scaling is available may be signaled at the slice level or tile group level.

For example, scaling can be applied only when mapping for luma components is available, and partitioning of luma components and partitioning of chroma components follow the same tree structure.

Scaling may be performed based on an average of values of samples of a luma prediction block corresponding to the chroma prediction block. In this case, when the target block uses inter prediction, the luma prediction block may mean a mapped luma prediction block.

A value required for scaling may be derived by referring to a look-up table using an index of a piece to which an average of values of samples of the luma prediction block belongs.

By performing scaling on the residual block using the finally derived value, the residual block may be converted into an inversely mapped region. Then, for the chroma component block, reconstruction, intra prediction, inter prediction, in-loop filtering, and reference picture storage may be performed in the inversely mapped region.

For example, information indicating whether mapping and/or inverse mapping of the luma component and chroma component is available may be signaled through a sequence parameter set.

A prediction block of the target block may be generated based on the block vector. A block vector may indicate displacement between a target block and a reference block. A reference block may be a block within a target image.

In this way, a prediction mode in which a prediction block is generated by referring to a target image may be referred to as an intra block copy (IBC) mode.

IBC mode can be applied to CUs of a specified size. For example, IBC mode can be applied to MxN CUs. Here, M and N may be 64 or less.

The IBC mode may include a skip mode, a merge mode, and an AMVP mode. In the case of skip mode or merge mode, a merge candidate list may be constructed, and a merge index may be signaled so that one merge candidate among merge candidates of the merge candidate list may be specified. A block vector of a specified merge candidate may be used as a block vector of a target block.

In case of AMVP mode, a differential block vector may be signaled. Also, the predicted block vector may be derived from the left neighboring block and the top neighboring block of the target block. Also, an index of which neighboring block is to be used may be signaled.

The prediction block of the IBC mode may be included in the target CTU or the left CTU, and may be limited to a block within a pre-constructed region. For example, the value of the block vector may be limited so that the prediction block of the target block is located within a specified region. The specified area may be an area of three 64x64 blocks that are encoded and/or decoded prior to the 64x64 block including the target block. As the value of the block vector is limited in this way, memory consumption and device complexity according to the implementation of the IBC mode can be reduced.

16 is a structural diagram of an encoding device according to an embodiment.

The encoding device 1600 may correspond to the aforementioned encoding device 100.

The encoding device 1600 includes a processing unit 1610, a memory 1630, a user interface (UI) input device 1650, a UI output device 1660, and storage that communicate with each other through a bus 1690. (1640). Also, the encoding device 1600 may further include a communication unit 1620 connected to the network 1699.

The processor 1610 may be a semiconductor device that executes processing instructions stored in a central processing unit (CPU), memory 1630, or storage 1640. The processing unit 1610 may be at least one hardware processor.

The processing unit 1610 may generate and process signals, data, or information input to the encoding device 1600, output from the encoding device 1600, or used inside the encoding device 1600, signals, Inspection, comparison, and judgment related to data or information can be performed. In other words, in an embodiment, data or information generation and processing, and data or information related inspection, comparison, and judgment may be performed by the processing unit 1610 .

The processing unit 1610 includes an inter prediction unit 110, an intra prediction unit 120, a switch 115, a subtractor 125, a transform unit 130, a quantization unit 140, an entropy encoding unit 150, and an inverse quantization unit. It may include a unit 160, an inverse transform unit 170, an adder 175, a filter unit 180, and a reference picture buffer 190.

An inter prediction unit 110, an intra prediction unit 120, a switch 115, a subtractor 125, a transform unit 130, a quantization unit 140, an entropy encoding unit 150, an inverse quantization unit 160, At least some of the inverse transform unit 170, the adder 175, the filter unit 180, and the reference picture buffer 190 may be program modules and may communicate with an external device or system. Program modules may be included in the encoding device 1600 in the form of an operating system, application program modules, and other program modules.

Program modules may be physically stored on various known storage devices. Also, at least some of these program modules may be stored in a remote storage device capable of communicating with the encoding device 1600 .

Program modules include routines, subroutines, programs, objects, components, and data that perform functions or operations according to an embodiment or implement abstract data types according to an embodiment. A data structure, etc. may be encompassed, but is not limited thereto.

Program modules may include instructions or codes executed by at least one processor of the encoding device 1600 .

The processing unit 1610 includes an inter prediction unit 110, an intra prediction unit 120, a switch 115, a subtractor 125, a transform unit 130, a quantization unit 140, an entropy encoding unit 150, and an inverse quantization unit. Instructions or codes of the unit 160, the inverse transform unit 170, the adder 175, the filter unit 180, and the reference picture buffer 190 may be executed.

Storage may represent memory 1630 and/or storage 1640 . Memory 1630 and storage 1640 may be various forms of volatile or non-volatile storage media. For example, the memory 1630 may include at least one of a ROM 1631 and a RAM 1632 .

The storage unit may store data or information used for the operation of the encoding device 1600 . In an embodiment, data or information of the encoding device 1600 may be stored in a storage unit.

For example, the storage unit may store pictures, blocks, lists, motion information, inter prediction information, and bitstreams.

The encoding device 1600 may be implemented in a computer system including a recording medium that can be read by a computer.

The recording medium may store at least one module required for the encoding device 1600 to operate. The memory 1630 may store at least one module, and the at least one module may be configured to be executed by the processing unit 1610 .

A function related to communication of data or information of the encoding device 1600 may be performed through the communication unit 1620 .

For example, the communication unit 1620 may transmit a bitstream to a decoding apparatus 1700 to be described later.

17 is a structural diagram of a decryption device according to an embodiment.

The decoding device 1700 may correspond to the decoding device 200 described above.

The decryption device 1700 includes a processing unit 1710, a memory 1730, a user interface (UI) input device 1750, a UI output device 1760, and storage that communicate with each other through a bus 1790. (1740). Also, the decoding apparatus 1700 may further include a communication unit 1720 connected to the network 1799.

The processor 1710 may be a semiconductor device that executes processing instructions stored in a central processing unit (CPU), memory 1730, or storage 1740. The processing unit 1710 may be at least one hardware processor.

The processing unit 1710 may generate and process signals, data, or information input to the decoding device 1700, output from the decoding device 1700, or used inside the decoding device 1700, and signals, Inspection, comparison, and judgment related to data or information can be performed. In other words, in an embodiment, data or information generation and processing, and data or information related inspection, comparison, and judgment may be performed by the processing unit 1710 .

The processing unit 1710 includes an entropy decoding unit 210, an inverse quantization unit 220, an inverse transform unit 230, an intra prediction unit 240, an inter prediction unit 250, a switch 245, an adder 255, a filter A unit 260 and a reference picture buffer 270 may be included.

An entropy decoding unit 210, an inverse quantization unit 220, an inverse transform unit 230, an intra prediction unit 240, an inter prediction unit 250, a switch 245, an adder 255, a filter unit 260, and At least some of the reference picture buffers 270 may be program modules, and may communicate with an external device or system. Program modules may be included in the decryption device 1700 in the form of an operating system, application program modules, and other program modules.

Program modules may be physically stored on various known storage devices. Also, at least some of these program modules may be stored in a remote storage device capable of communicating with the decryption device 1700 .

Program modules include routines, subroutines, programs, objects, components, and data that perform functions or operations according to an embodiment or implement abstract data types according to an embodiment. A data structure, etc. may be encompassed, but is not limited thereto.

The program modules may include instructions or codes executed by at least one processor of the decoding device 1700 .

The processing unit 1710 includes an entropy decoding unit 210, an inverse quantization unit 220, an inverse transform unit 230, an intra prediction unit 240, an inter prediction unit 250, a switch 245, an adder 255, a filter Instructions or codes of unit 260 and reference picture buffer 270 may be executed.

Storage may represent memory 1730 and/or storage 1740 . Memory 1730 and storage 1740 may be various forms of volatile or non-volatile storage media. For example, the memory 1730 may include at least one of a ROM 1731 and a RAM 1732 .

The storage unit may store data or information used for the operation of the decoding device 1700 . In an embodiment, data or information of the decoding device 1700 may be stored in the storage unit.

For example, the storage unit may store pictures, blocks, lists, motion information, inter prediction information, and bitstreams.

The decryption device 1700 may be implemented in a computer system including a recording medium that can be read by a computer.

The recording medium may store at least one module required for the decoding apparatus 1700 to operate. The memory 1730 may store at least one module, and the at least one module may be configured to be executed by the processing unit 1710 .

A function related to communication of data or information of the decoding device 1700 may be performed through the communication unit 1720 .

For example, the communication unit 1720 may receive a bitstream from the encoding device 1600.

Hereinafter, the processing unit may represent the processing unit 1610 of the encoding device 1600 and/or the processing unit 1710 of the decoding device 1700. For example, for functions related to prediction, the processing unit may represent switch 115 and/or switch 245. In the function related to inter prediction, the processing unit may represent the inter prediction unit 110, the subtractor 125, and the adder 175, and may represent the inter prediction unit 250 and the adder 255. In the function related to intra prediction, the processing unit may represent the intra prediction unit 120, the subtractor 125, and the adder 175, and may represent the intra prediction unit 240 and the adder 255. In the function related to transform, the processing unit may represent the transform unit 130 and the inverse transform unit 170, and may represent the inverse transform unit 230. In the function related to quantization, the processing unit may represent the quantization unit 140 and the inverse quantization unit 160, and may represent the inverse quantization unit 220. In the function related to entropy encoding and/or decoding, the processing unit may represent the entropy encoding unit 150 and/or the entropy decoding unit 210. In terms of filtering-related functions, the processing unit may represent the filter unit 180 and/or the filter unit 260. For functions related to reference pictures, the processing unit may indicate the reference picture buffer 190 and/or the reference picture buffer 270 .

Embodiments for lightening an in-loop filter and prediction deep neural network

As deep neural network technology develops, techniques based on deep neural networks are being proposed even in video compression. Video compression techniques based on deep neural networks can show superior performance than conventional video compression techniques. On the other hand, video compression techniques based on deep neural networks may have the disadvantage of requiring a very large execution time compared to conventional video compression techniques due to the enormous amount of computation of deep neural networks.

In the following embodiments, a method, apparatus, and recording medium for efficiently lightening an in-loop filter and a predictive deep neural network may be provided in order to improve encoding efficiency of an image and reduce a time required for encoding.

In the following embodiments, a method, apparatus, and recording medium for providing learning in a lightweight in-loop filter and lightweight predictive deep neural network through a knowledge distillation technique may be provided. there is.

Through the following embodiments, an in-loop filter for improving the efficiency and speed of encoding an image and a learning method in a predictive deep neural network may be provided.

In embodiments, the terms “deep neural network” and “neural network” may be used interchangeably. That is to say, “deep” is an optional expression and may be removed or omitted in embodiments.

In embodiments, the terms “neural network” and “network” may be used interchangeably. That is to say, the term "neural network" and the term "network" may have the same meaning.

In embodiments, the terms “learning” and “training” may be used interchangeably. That is to say, the term "learning" and the term "training" can have the same meaning.

18 is a flowchart of an encoding method according to an embodiment.

A deep neural network may include one or more of an in-loop filter and a predictive deep neural network. Alternatively, the deep neural network may represent one or more of an in-loop filter based on the deep neural network and a predictive deep neural network for prediction. In embodiments, the deep neural network may be replaced with an in-loop filter and/or predictive deep neural network.

In step 1810, the processing unit 1610 of the encoding device 1600 may perform learning in the deep neural network.

At step 1820, processing unit 1610 may select a deep neural network.

In step 1830, the processing unit 1610 may perform encoding on deep neural network information. The processing unit 1610 may generate encoded deep neural network information by encoding the deep neural network information.

In operation 1840, the processing unit 1610 may generate a bitstream including deep neural network information or encoded deep neural network information.

The processing unit 1610 may store the bitstream in the storage unit 1630 .

The communication unit 1620 may transmit the bitstream to the decoding device 1700.

Deep neural network information or encoded deep neural network information may be signaled from the encoding device 1600 to the decoding device 1700 through a bitstream.

19 is a flowchart of a decoding method according to an embodiment.

A deep neural network may include one or more of an in-loop filter and a predictive deep neural network. Alternatively, the deep neural network may represent one or more of an in-loop filter based on the deep neural network and a predictive deep neural network for prediction. In embodiments, the deep neural network may be replaced with an in-loop filter and/or predictive deep neural network.

In step 1910, the communication unit 1730 may receive a bitstream from the encoding device 1600.

In step 1920, the processing unit 1710 of the decoding device 1700 may perform learning in the deep neural network.

Steps 1910 and 1920 may be performed concurrently. Alternatively, step 1920 may be performed before step 1910 .

The storage unit 1720 may store a bitstream.

The processing unit 1710 may obtain a bitstream from the storage unit 1720 or the communication unit 1730.

The bitstream may include deep neural network information or encoded deep neural network information.

In step 1930, the processor 1610 may decode the encoded deep neural network information. The processing unit 1610 may generate deep neural network information by encoding the encoded deep neural network information.

At step 1940, processing unit 1610 may select a deep neural network.

20 is a flowchart of learning in a deep neural network according to an example.

Step 1810 may include steps 2010, 2020 and 2030 below.

Step 1920 may include steps 2010, 2020 and 2030 below.

In embodiments, learning in a deep neural network may refer to learning in an in-loop filter and/or learning in a predictive deep neural network.

At step 2010, the processing unit may perform learning on a teacher network.

At step 2020, the processing unit may perform a first learning in the student network. The first learning may refer to first stage learning.

At step 2030, the processing unit may perform a second learning in the student network. The second learning may refer to second stage learning.

In embodiments, the size of the input of the in-loop filter may be equal to the size of the reconstructed image or reconstructed block (before in-loop filtering is applied).

Learning in the filter described in the embodiments may be learning under a specific condition or learning for a specific target.

The specific condition may be a condition in which the value of a coding parameter is a specific value. For example, a specific condition may be that a value of a quantization parameter is a specific value. For example, the specific value could be 22, 27, 32, 37 or 42.

A specific object may be an image or a component of a block. For example, a specific object may be a luma component and/or a chroma component.

A specific target may be a slice type. _For example, slice types may include Intra _A , Intra _B , and Inter _.

Due to encoding QP fluctuations, low-latency B and Random Access (RA) can utilize intra-slice filters trained using different quantization parameters. For example, different quantization parameters may be {19, 24, 29, 34, 39}.

In embodiments, the in-loop filter may restore a reconstructed image or a reconstructed block for each component of an input image.

21 illustrates learning in a teacher network according to an example.

In step 2010, in-loop filter learning based on a deep neural network may be performed to improve the quality of an image in which compression deterioration has occurred.

For example, learning in an in-loop filter based on a deep neural network may be performed as shown in FIG. 21 .

An in-loop filter may include a filtering deep neural network. A filtering deep neural network may include a plurality of modules.

An input of the in-loop filter (or deep neural network filtering) may include an image in which compression degradation occurs and coding information of the image. An image in which compression deterioration has occurred may be input to a first module among a plurality of modules. Coding information of an image may be input to each module of a plurality of modules. Coding information of an image may be input to one or more modules among a plurality of modules.

An output of the in-loop filter (or deep neural network filtering) may be an image from which compression degradation has been removed.

In embodiments, elimination of compression degradation may not imply complete elimination of compression degradation. For example, an image from which compression degradation is removed may be an image from which compression degradation is partially removed.

For example, an image from which compression degradation has been removed may be an image generated by applying a specific process to an image in which compression degradation occurs.

For example, an image from which compression degradation has been removed may be an image closer to an original image than an image in which compression degradation has occurred.

Also, an image from which compression degradation has been removed may simply mean an output of an in-loop filter, and the in-loop filter may perform processing to remove or reduce compression degradation in the image.

In embodiments, an image may refer to a target image including a target block.

In embodiments, the coding information may include the aforementioned coding parameters. In other words, the coding information may be information related to the target image or target block described in the embodiments.

For example, the input of the in-loop filter may be an image in which compression degradation occurs.

For example, the output of the in-loop filter may be an image from which compression degradation has been removed.

For example, learning in an in-loop filter based on a deep neural network may be performed so that an error between an output of the in-loop filter and an original image is minimized.

In embodiments, an original image may refer to an original block in an original image corresponding to a target block.

For example, learnings in a plurality of in-loop filters may be performed according to various degrees of compression degradation.

For example, an in-loop filter based on a deep neural network to be trained may be an image reconstruction deep neural network.

For example, the image reconstruction deep neural network may include a plurality of convolutional neural network layers or fully-connected neural network layers.

For example, a plurality of neural network layers may be grouped into one module.

22 shows learnings in an intra-prediction deep neural network and an inter-prediction deep neural network according to an embodiment.

A deep neural network can be a predictive deep neural network. The predictive deep neural network may generate a prediction image (or prediction block) similar to the original image (or original block) through prediction.

In step 2010, learning in the predictive deep neural network may be performed to generate a prediction image (or prediction block) similar to the original image (or original block) through intra prediction and/or inter prediction.

In embodiments, the term “block” may be replaced with the term “signal”. For example, the term “prediction block” may be replaced with the term “prediction signal”.

For example, learning in a predictive deep neural network can be configured as shown in FIG. 22 .

The predictive deep neural network may include an intra-predictive deep neural network and/or an inter-predictive deep neural network. For example, the predictive deep neural network may be an intra-predictive deep neural network or an inter-predictive deep neural network.

An inter-prediction deep neural network may include a plurality of modules.

Inputs to the intra-prediction deep neural network may include neighboring samples of the target block and coding information of an image. The surrounding samples may be plural.

The output of the intra-prediction deep neural network may be an improved prediction block of the target block.

Inputs to the inter-prediction deep neural network may include a prediction block of a target block and coding information of an image. A prediction block of the target block may be input to a first module among a plurality of modules. Coding information of an image may be input to each module of a plurality of modules. Coding information of an image may be input to one or more modules among a plurality of modules.

The output of the inter-prediction deep neural network may be an improved prediction block of the target block.

Alternatively, an improved prediction block of the target block may be generated using the output of the intra-prediction deep neural network and the output of the inter-prediction deep neural network.

For example, an input of the intra prediction deep neural network may include at least one of a neighboring sample of the target block and a prediction block of the target block. The neighboring samples may be samples adjacent to the target block. The peripheral sample may be the reference sample described above in the embodiments. Neighboring samples may be samples that have already been encoded/decoded.

For example, learning in the predictive deep neural network may be performed to minimize an error between the output of the predictive deep neural network and the original image.

For example, an input of an inter-prediction deep neural network may be a prediction block of a current block. A prediction block may be a block in a reference picture. A prediction block may be determined through compensation using motion that minimizes rate-distortion loss. It may be a prediction block in . A reference picture may be a frame that has already been encoded/decoded.

For example, learning in a plurality of predictive deep neural networks may be performed according to various degrees of compression degradation.

For example, a predictive deep neural network may include multiple convolutional neural network layers or fully-connected neural network layers.

For example, a plurality of neural network layers may be grouped into one module.

Learning from teacher networks

In the above step 2010, the coding information of the image may be used for learning in the teacher network.

For example, as in the embodiments described above with reference to FIGS. 21 and 22 , coding information of an image used in learning in a teacher network may include a coding parameter. For example, the coding information of an image includes 1) information on block division of the image, 2) information on the encoding mode of the target block, 3) information on the encoding mode of neighboring blocks, and 4) information on quantization of the target block. and 5) information on the number of times the in-loop filter is performed on the target block.

For example, coding information of an image may be input to a teacher network. At this time, the coding information of the image may be input to the teacher network together with the input image of the teacher network or neighboring samples.

For example, coding information of an image may be input to a specific module of the teacher network. At this time, coding information of the image may be input to a specific module together with an output of a previous module. A previous module can be a module linked to a specific module. The previous module may be a module located before a specific module.

For example, coding information of an image may be converted according to a specific position corresponding to an output of a specific module of the teacher network, and the converted coding information may be input to the specific position.

For example, encoding information of an image may be input to directly change model parameters within a module of a teacher network.

The trained deep neural network described in the embodiments can be utilized as a teacher network in knowledge distillation.

Additional information may be input to the above-described filtering deep neural network and prediction deep neural network. The additional information may include a prediction image (or prediction block), a slice quantization parameter, a base quantization parameter, and a slice type. Also, the additional information may include coding parameters.

First Learning in Student Network

In step 2020, the student network may be constructed with a deep neural network structure similar to that of the teacher network.

In embodiments, the pre-trained teacher network may be a deep neural network with greater complexity compared to the student network on which learning is to be performed. Meanwhile, the complexity of the teacher network and the complexity of the student network may be the same. The difference between the complexity of the teacher network and the complexity of the student network may not be large.

For example, the number of modules in the teacher network and the number of modules in the student network may be the same. Here, the structure of the modules of the teacher network and the structure of the modules of the student network may be the same or similar to each other.

For example, the size of the feature map of the teacher network and the size of the feature map of the student network may be the same.

For example, the size of the activity vector of the teacher network and the size of the activity vector of the student network may be the same.

Learning in a student network can be performed using knowledge distillation using a pre-trained teacher network. When learning using knowledge distillation is performed, information of the teacher network may be transferred to the student network, and learning may be performed in the student network using the transferred information.

Through learning in the teacher network, the teacher network can be pre-trained. Next, the student network can be trained using Attention Transfer (AT) loss. Intermediate features of the teacher network and the student network can be averaged along the channel axis to obtain one-channel attention maps. In knowledge distillation, a student network can be trained to mimic the attention map obtained from the teacher network. For effective knowledge distillation, features can be mapped to a map of states.

For example, the degree of compression degradation of the input of the teacher network and the degree of compression degradation of the input of the student network may be the same.

For example, the size of neighboring samples input to the teacher network and the size of neighboring samples input to the student network may be the same.

For example, learning in an in-loop filter based on a deep neural network may be performed until knowledge distillation errors converge through knowledge distillation.

For example, when learning using knowledge distillation is performed, additional learning may not be performed in the learned teacher network, and learning may be performed only in the student network. In embodiments, learning may refer to learning within a student network.

For example, when learning using knowledge distillation is performed, learning may be performed in both the teacher network and the student network together. In embodiments, learning may refer to learning within a teacher network and learning within a student network.

For example, learning may be performed such that values of corresponding model parameters in the deep neural network of the teacher network and the deep neural network of the student network are equal.

For example, learning may be performed such that an error between the result of the teacher network and the result of the student network is minimized.

For example, learning on student networks can be performed using intermediate feature maps or activation vectors. The above intermediate feature map or activation vector may be an output of a specific module among a plurality of modules of the teacher network.

For example, a first module in the teacher network and a second module in the student network may correspond to each other. Correspondence between the first module in the teacher network and the second module in the student network may be established. Learning may be performed so that outputs from modules corresponding to each other in the teacher network and the student network have characteristics similar to each other. Here, the outputs may include intermediate feature maps or activity vectors.

For example, learning may be performed on the teacher network and the student network such that it is difficult for a separate discriminator deep neural network to distinguish between outputs from the teacher network and outputs from the student network. Here, the output may include a feature map, an active vector, or an output image. The output image may be an image from which compression deterioration has been removed. For example, feature maps or active vectors may be extracted from modules at the same position among the entire structure of the teacher network and the entire structure of the student network. Learning may be performed using extracted feature maps or extracted learning vectors.

For example, results from all modules of the teacher network and the student network (eg, feature maps or activity vectors) may be used for learning.

For example, training may be performed such that the complexity of the student network is minimized. Here, learning can be performed such that the complexity of the student network and the knowledge distillation error are minimized at the same time.

For example, the information of the teacher network may include one or more of model parameters, intermediate feature maps, active vectors, and filtered images. Intermediate feature maps and activation vectors can be functions of specific training images. The learning image may be an input image input to the teacher network. The filtered image may be the (final) output from the teacher network.

For example, information of a plurality of teacher networks may be used for learning in the teacher network and the student network.

For example, information of a specific teacher network among a plurality of teacher networks may be used for learning. Here, the specific teacher network may be a teacher network that generates an output having the smallest error with respect to the original video among a plurality of teacher networks.

For example, the information of the teacher network may include coding information of a learning image. Information of the teacher network including coding information of the learning image may be used for learning.

Learning using knowledge distillation in embodiments may include one or more of the approaches described above.

In relation to the attention map, the calculation process of the attention module can be expressed as Equation 1 below.

[Equation 1]

F_out = F_in * f ( Rec , Pred , BS , QP ) + F_in

F_in may be an input of an attention module. F_out may be the output of the attention module.

Rec, Pred, BS, and QP may represent reconstructed images, predicted images, boundary strengths, and quantization parameters, respectively. The quantization parameter may be a sequence-level input quantization parameter. f may consist of two convolutions. The activation function can be applied after the first convolutional layer. The purpose of f may be to generate an attention map from external information, and then to perform recalibration on the feature map F_in .

Inputs to the neural network may include reconstructed images, predicted images, boundary strengths, and quantization parameters. (Split can additionally be taken as an input within intra-slice models.)

Whether or not filtering is to be applied may be determined for each slice or each block. When filtering is determined to be applied to a slice or block, a conditional parameter may be additionally determined from the candidate list. The candidate list may include three candidates derived from the quantization parameter.

q may represent a sequence level quantization parameter. The candidate list may include conditional parameters {q, q-5, q-10}. This selection process may be based on the rate-distortion cost of the encoding device 1600.

The deep neural network information may include information indicating whether to use the selection process and indexes for conditional eye parameters. The index may indicate a conditional parameter used for filtering among conditional parameters.

The granularity of the decision of the filter and the choice of parameters may vary based on the resolution and quantization parameters. For example, given higher resolution and larger quantization parameters, decisions and selections can be performed over larger domains.

For different temporal layers, the candidate parameter list for the temporal layer may be different. For high temporal layers, the third candidate q - 10 can be replaced by q + 5.

For an all-intra configuration, parameter selection can be disabled while on/off control is still maintained.

When a neural network filter is applied to a reconstructed image, a scaling factor may be signaled for each color component in a picture header. The difference between the input image and the image filtered by the neural network may be scaled by a scaling factor before the difference is added to the input image. This scaling may be termed residual scaling.

To enable combination with deblocking, the input image used in residual scaling may be the output of a deblocking filter.

Equation 2 below may represent a process of residual scaling.

[Equation 2]

R _Refine = ( R _NN - R _DB ) * w + R _DB = w * R _NN + (1 - w ) * R _DB

R _NN can be the output of a neural network filter.

R _DB can be the output of the deblocking filter.

w may be a weight.

R _Refine can be the result of residual scaling.

23 illustrates a first learning in a student network according to one embodiment.

In step 2020, learning in a learning network using a teacher network may be performed for learning in an in-loop filter using knowledge distillation.

For example, first learning in a student network using knowledge distillation may be performed according to the configuration shown in FIG. 23 .

For example, feature maps that are results of a module of a teacher network and a module of a student network can be extracted. Each feature map of the extracted feature maps may be converted into an attention map. Attention maps may be created using the extracted feature maps. Here, an average may be derived for each axis of the feature map, and an attention map may be constructed using the derived average. A sum may be derived for each axis of the feature map, and an attention map may be constructed using the derived sum.

For example, the first feature map may be the result of a module of a teacher network. The second feature map may be the result of a module of the student network. When the size of the first feature map and the size of the second feature map are different, the attention map may be constructed by transforming the feature map using a transform module having learnable parameters.

For example, the first attention map may be an attention map obtained from a result of a module of the teacher network. The second attention map may be an attention map obtained from the result of the module of the student network. Learning may be performed to minimize an error between the first attentional map and the second attentional map.

Here, the learning loss function may be configured as in Equation 3 below.

[Formula 3]

L _step1 may represent a loss function of learning.

may represent the attention map derived from the feature map output from the ith module of the teacher network.

may represent the attention map derived from the feature map output from the i-th module of the student network.

24 illustrates a first learning in a student network according to one embodiment.

In step 2020, for learning in a predictive deep neural network using knowledge distillation, learning in a learning network using a teacher network may be performed.

For example, first learning in a student network using knowledge distillation may be performed according to the configuration shown in FIG. 24 .

For example, activation vectors that are results of a module of a teacher network and a module of a student network can be extracted. The first activation vector may be the result of a module of the teacher network. The second activation vector may be the result of a module of the student network. A first activation vector can be extracted from the module of the teacher network. A second activation vector may be extracted from a module of the student network. Here, a transformation module for equalizing the dimensions of the first activation vector and the dimensions of the second activation vector may be used. A plurality of transform modules may be added to the predictive deep neural network.

For example, the transform module may change the dimension of the second active vector to be the same as the dimension of the first active vector by performing transform on the second active vector. The transform module may generate a second active vector having a changed dimension by performing a transform on the second active vector. Parameters of the transformation module may be determined by learning.

For example, learning may be performed such that an error between activation vectors from a module of a teacher network and a module of a student network is minimized. Here, the active vectors may be a first active vector and a second active vector. Alternatively, the activity vectors can be a first activity vector and a second activity vector with altered dimensions.

Here, the learning loss function may be configured as in Equation 4 below.

[Formula 4]

L _step1 may represent a loss function of learning.

may represent the first activation vector output from the i-th module of the teacher network.

may be a second activation vector having a modified dimension generated by performing transformation on a second activation vector output from the i-th module of the student network.

25 illustrates second learning in a student network according to one embodiment.

In step 2030, second learning may proceed with respect to the student network that has completed the first learning. In the second learning, 1) the original image or surrounding samples and 2) the output of the teacher network may be used.

In the second learning, fine-tune may be performed on the student network to which the first learning is applied. Here, the output of the teacher network and the original image can be used as a supervision signal for fine-tuning. Both directors can be used to measure the L1 loss. Each of the two directors can be named imitation loss (to the output of the teacher network) and reconstruction loss (to the original video) respectively.

The purpose of the second learning can be expressed as Equation 5 below.

[Formula 5]

L _Finetune = L _{Reconstuction} + α * L _imitation

L _Finetune can be an L1 loss. L _{Reconstuction} may be the reconstruction loss. L _imitation may be imitation loss.

The effect of the two loss terms can be adjusted using the hyperparameter α. α can be (empirically) set to 1.

For example, the second learning of the student network may be performed according to the configuration shown in FIG. 25 .

For example, training in an in-loop filter and predictive deep neural network may be performed such that the error between the output of the student network and the output of the teacher network is minimized.

In embodiments, the output of the student network may be a resulting image output from the student network. The output of the teacher network may be a result image output from the teacher network.

For example, training in an in-loop filter and predictive deep neural network may be performed so that an error between the output of the student network and the original image is minimized.

For example, learning in the in-loop filter and predictive deep neural network may be performed such that the first error and the second error are simultaneously minimized. The first error may be an error between the output of the student network and the output of the teacher network. The second error may be an error between the output of the student network and the original image. Here, weights of the first error and the second error in learning may be adjusted by a Lagrange multiplier. For example, the value of the Lagrange multiplier may be set to 1.

For example, learning in the in-loop filter and prediction deep neural network may be performed until the first error and the second error converge. Here, the learning loss function may be configured as in Equation 6 below.

[Formula 6]

L _step2 may represent a loss function of learning.

may represent the first error.

may represent a second error.

For example, the first loss function of the first learning in the student network and the second loss function of the second learning in the student network can be summed as one third loss function. Learning in the in-loop filter and predictive deep neural network may be performed using the third loss function.

Choice of deep neural networks

At steps 1820 and 1940, a deep neural network may be selected. Selecting a deep neural network may mean performing filtering on a target image using the selected deep neural network. Here, the target image may be a reconstructed image. The target image may mean a target block or a reconstructed block.

In steps 1820 and 1940, filtering of the target image may be performed using a filter based on the learned deep neural network. The deep neural network filter may be a trained deep neural network described in the embodiments. Alternatively, the deep neural network filter may be a filter using a trained deep neural network described in the embodiments.

In embodiments, a filter may refer to a deep neural network filter.

When filtering is performed on the target image, one filter out of N filters may be selected. Here, the N filters may be N deep neural network filters. The N deep neural network filters may each perform image reconstructions of different degrees. That is to say, the N deep neural network filters may differ from each other in the degree of image reconstruction.

For example, the N deep neural network filters may be student networks trained using knowledge distillation.

For example, whether or not a filter is applied to a specific target may be determined. For example, a specific object may be a unit including a target block.

In embodiments, a specific object is a (reconstructed) picture, (reconstructed) picture, (reconstructed) slice, (reconstructed) Coding Tree Unit (CTU), or (reconstructed) may be a block.

Here, the size of the (reconstructed) block may be HxW. N, H and W may be positive integers.

For example, when filtering is performed on a specific target, one filter out of N filters may be selected.

For example, a filter candidate set may be configured according to the compression degree and compression method for a specific object (or image). A filter candidate set may be some filters among N filters. In the filter selection described in the embodiments, one filter may be selected from among the filters of the filter candidate set. A filter candidate set may be M deep neural network filters. M may be an integer less than or equal to N. The M deep neural network filters of the filter candidate set may each perform image reconstructions of different degrees.

For example, a filter having the smallest error among N filters or filters of a filter candidate set may be selected. Here, the error may be an error between the result image of the filter and the original image.

For example, a filter may be selected using the result of filtering on the reference unit. For example, the reference unit may include neighboring blocks, CTUs, slices, and reference images adjacent to the target image.

In embodiments, the filter's weight α may be used for the filter.

For example, one weight may be used and different weights may be used according to the characteristics of the target image. The target image may be an input image that is a compression target.

For example, deep neural network information may include weight information. The weight information may indicate weight α. Weight information may be signaled through a bitstream.

For example, the number of variable weights may be determined based on characteristics of the target image. The number of weights may be determined in terms of rate-distortion in an image compression process.

For example, the output of the filter may be adjusted by using a weight to minimize an error between the output of the filter and the original image. Equation 7 below may indicate an example of adjusting the output of the filter using weights.

[Formula 7]

I _input may be an original video. I _filtered can be the output of the filter. I _output may be a reconstructed image. I _output may be a reconstructed image to which an (in-loop) filter is applied. α may be a weight.

For example, a weighted sum of an output of a deep neural network filter and a result of an existing in-loop filter may be used as a reconstructed image. At this time, one or more results among the results of the existing in-loop filter may be used for the weighted sum. A result of the in-loop filter may mean a reconstructed image to which the in-loop filter is applied. Equation 8 below may represent a weighted sum using the output of the deep neural network filter.

[Formula 8]

I _CNN can be the output of a deep neural network filter. I _LF may be the result of an in-loop filter. I _output may be a reconstructed image. I _output may be a reconstructed image to which an (in-loop) filter is applied. α may be a weight.

For example, a weight used for output of a deep neural network filter for a specific target may be adjusted. For example, a specific object may be a unit including a target block.

In embodiments, a specific object may be a (reconstructed) picture, (reconstructed) picture, (reconstructed) slice, (reconstructed) (CTU) or (reconstructed) block.

For example, for a particular subject, a weighted sum of the output of a deep neural network filter and the result of an existing in-loop filter can be used.

For example, for a specific target, the weight for the weighted sum of the output of the deep neural network filter and the result of the existing in-loop filter may be adjusted.

For example, for a specific target, one more suitable in terms of rate-distortion among the output of the deep neural network filter and the result of the existing in-loop filter may be selected as the reconstructed image (with the in-loop filter applied). there is.

For example, the results of an in-loop filter may be multiple. Among the plurality of results of the in-loop filter, one more suitable in terms of rate-distortion may be selected as the reconstructed image (to which the in-loop filter is applied).

For example, one or more results of a plurality of results of an existing in-loop filter may be used as candidates in the aforementioned selection.

In embodiments, existing in-loop filters include a deblocking filter, a Sample Adaptive Offset (SAO), an Adaptive Loop Filter (ALF) and a cross-component adaptive loop filter ( Cross Component Adaptive Loop Filter (CCALF) and the like. However, the in-loop filters of the embodiments are not limited to deblocking filters, SAO, ALF and CCALF, and the like. Existing in-loop filters may further include the in-loop filters of the embodiments and other in-loop filters.

For example, a deep neural network filter may be applied before an existing in-loop filter. A deep neural network filter may be applied to a specific target before an existing in-loop filter is applied to the specific target. Deep neural network filters can be applied later than conventional in-loop filters. A deep neural network filter may be applied to a specific target after a conventional in-loop filter is applied to the specific target. A deep neural network filter may be applied between existing in-loop filters. The deep neural network filter may be applied to the specific object after the first conventional in-loop filter is applied to the specific target, and the second conventional in-loop filter may be applied to the specific target after the deep neural network filter is applied to the specific target. It can be applied to specific targets.

For example, coding information of an image may be used in selecting a deep neural network filter.

In embodiments, one or more of the methods for selecting a deep neural network filter described above may be combined.

Filtering of the reconstructed image using the determined deep neural network filter

A deep neural network filter to be applied to the reconstructed image may be determined through the selection of the aforementioned deep neural network filter. In an embodiment, a "selected filter" may mean a "determined filter".

Hereinafter, a deep neural network may refer to a deep neural network filter selected/determined to be applied to a reconstructed image.

Hereinafter, the reconstructed image may be replaced with the specific unit described above. For example, a “reconstructed image” may be considered a (reconstructed) image, a (reconstructed) slice, a (reconstructed) CTU, or a (reconstructed) block.

Filtering may be performed on the reconstructed image using the determined deep neural network filter.

For example, the input of a deep neural network filter may be a reconstructed image generated in a loop. The input of the deep neural network filter may include the reconstructed image generated within the loop.

For example, an input of a deep neural network filter may be a luma component and/or a chroma component of a reconstructed image generated within a loop. An input of the deep neural network filter may include a luma component and/or a chroma component of the reconstructed image generated within the loop.

In embodiments, the size of the luma component may be greater than the size of the chroma component. Up-sampling of the chroma component may be performed to match the size of the chroma component to the size of the luma component. When a chroma component is input to a deep neural network filter or a predictive deep neural network, up-sampling may be performed on the chroma component. Up-sampling of chroma components may be performed as pre-processing for filtering or prediction by a deep neural network filter by a deep neural network. If the chroma component is output from the deep neural network filter or prediction deep neural network, down-sampling of the chroma component may be performed. Down-sampling of chroma components may be performed as pre-processing for filtering or prediction by a deep neural network filter by a deep neural network.

In embodiments, the chroma component may be a Cb component and/or a Cr component.

For example, when filtering is performed on a chroma component of a reconstructed image by a deep neural network filter, the luma component of the reconstructed image together with the chroma component of the reconstructed image may be input to the deep neural network filter.

For example, when filtering is performed on a chroma component of a reconstructed image by a deep neural network filter, the luma component of the reconstructed image may be included as an input of the deep neural network filter.

In embodiments, the luma component may be one of a predicted luma component (ie, the luma component of the predicted image), a reconstructed luma component (ie, the luma component of the reconstructed image), and a filtered reconstructed luma component. there is. The filtered reconstructed luma component may be a luma component (of the reconstructed image) to which in-loop filtering by a deep neural network filter or an existing in-loop filter is applied.

For example, the input of a deep neural network filter may be a filtered reconstructed image generated in a loop. A filtered reconstructed image may be generated by applying (in-loop) filtering to the reconstructed image. The input of the deep neural network filter may include the filtered reconstructed image generated within the loop.

For example, the deep neural network information may include index information indicating a filter selected in the encoding device 1600. Index information may be signaled through a bitstream. The processing unit 1710 of the decoding apparatus 1700 may identify a filter selected from among a plurality of filters using the index information, and perform filtering on the reconstructed image using the selected filter.

For example, the processing unit 1710 of the decoding apparatus 1700 may estimate or restore index information using other information such as deep neural network information and prediction information. The processing unit 1710 of the decoding apparatus 1700 may identify a selected filter indicated by the index information among a plurality of filters, and perform filtering on the reconstructed image using the selected filter. The processing unit 1710 of the decoding apparatus 1700 may identify a selected filter indicated by the index information among a plurality of filters, and may perform in-loop filtering using a deep neural network using the selected filter.

Prediction Using a Trained Prediction Deep Neural Network

Intra-prediction and/or inter-prediction may be performed using the aforementioned trained predictive deep neural network. When intra prediction and/or inter prediction are performed, one filter out of N filters may be selected.

N can be 1, 2 or 3. Alternatively, N may be a positive integer.

The N filters may be N predictive deep neural networks. N predictive deep neural networks may use neighboring samples of the target block as inputs to different degrees. That is to say, the N predictive deep neural networks can differ from each other in terms of the degree to which they use surrounding samples as input. For example, N predictive deep neural networks may use different numbers of surrounding samples as input.

For example, the N predictive deep neural networks may be student networks to which learning using knowledge distillation has been applied.

For example, whether or not to apply a filter may be determined for a specific target.

In embodiments, a specific object may include a video, picture, slice, CTU, and block.

Here, the size of the block may be HxW. H and W may be positive integers.

For example, when intra-prediction and/or inter-prediction are performed on a specific target, one of N predictive deep neural networks may be selected.

For example, a predictive deep neural network candidate set may be configured according to the compression degree and compression method for a specific object (or image). The predictive deep neural network candidate set may be some predictive deep neural networks among the N predictive deep neural networks. In the selection of the predictive deep neural network described in the embodiments, one predictive deep neural network may be selected from among the predictive deep neural networks of the set of predictive deep neural network candidates. The predictive deep neural network candidate set may be M predictive deep neural networks. M may be an integer less than or equal to N. The M predictive deep neural networks of the predictive deep neural network candidate set may each perform different predictions.

For example, a predictive deep neural network may be selected based on a size of a block from among N predictive deep neural networks or predictive deep neural networks of a predictive deep neural network candidate set.

For example, a prediction deep neural network having the smallest error among N prediction deep neural networks or prediction deep neural networks of a prediction deep neural network candidate set may be selected. Here, the error may be an error between the original image and the result of prediction by the predictive deep neural network (ie, an improved prediction block).

For example, a predictive deep neural network may be selected using the results of prediction or reconstruction on the reference unit. For example, the reference unit may include neighboring blocks, CTUs, slices, and reference images adjacent to the target image.

For example, weights for a predictive deep neural network may be changed based on characteristics of a target image. In other words, weights for the predictive deep neural network may be determined based on the characteristics of the target image. The target image may be an input image that is a compression target.

For example, deep neural network information may include weight information. Weight information may represent weights for predictive deep neural networks. Weight information may be signaled through a bitstream.

For example, the number of variable weights may be determined based on characteristics of the target image. The number of weights may be determined in terms of rate-distortion in an image compression process.

For example, a predictive deep neural network may be selected using coding information of an image.

In embodiments, one or more of the methods for selecting a predictive deep neural network described above may be combined.

Intra prediction and/or inter prediction may be performed using the determined predictive deep neural network.

For example, when intra prediction is performed, the input of the predictive deep neural network may be neighboring samples of the target block. When intra prediction is performed, the input of the predictive deep neural network may include neighboring samples of the target block.

For example, when inter prediction is performed, the input of the predictive deep neural network may be at least one of an image to which inter prediction is applied and neighboring samples of the target block. An input of the predictive deep neural network may include at least one of an image to which inter prediction is applied and neighboring samples of the target block.

For example, when intra prediction and/or inter prediction are performed, the input of the predictive deep neural network may be a luma component and/or a chroma component of an image. An input of the predictive deep neural network may include a luma component and/or a chroma component of an image.

For example, in intra prediction and/or inter prediction, when prediction of a chroma component of an image is performed by the predictive deep neural network, the chroma component of the image and the luma component of the image may be input to the predictive deep neural network. .Can be input to a deep neural network filter.

For example, in intra prediction and/or inter prediction, when prediction of a chroma component of an image is performed by the predictive deep neural network, the luma component of the image may be included as an input of the predictive deep neural network.

For example, the deep neural network information may include index information indicating a prediction deep neural network selected in the encoding device 1600 . Index information may be signaled through a bitstream. The processing unit 1710 of the decoding apparatus 1700 may identify a predicted deep neural network selected from among a plurality of predictive deep neural networks by using the index information, and may use the selected predictive deep neural network to intra-target blocks (or images). Prediction and/or inter prediction may be performed.

For example, the processing unit 1710 of the decoding apparatus 1700 may estimate or restore index information using other information such as deep neural network information and prediction information. A predictive deep neural network used to perform prediction can be specified using the index information. The processing unit 1710 of the decoding apparatus 1700 may identify a selected predictive deep neural network indicated by the index information among a plurality of predictive deep neural networks, and intra-predict a target block (or image) using the selected predictive deep neural network. and/or inter prediction.

For example, in intra prediction and/or inter prediction, when prediction of a chroma component of a target block (or image) is performed by a predictive deep neural network, the luma component of the target block (or image) is predicted by deep prediction. It can be included in the input of a neural network.

In embodiments, the luma component may include a predicted luma component (ie, a luma component of a predicted block or predicted image), a reconstructed luma component (ie, a luma component of a reconstructed block or reconstructed image), and a filtered reconstructed luma component. It can be one of the built luma components. The filtered reconstructed luma component may be a luma component (of a reconstructed block or a reconstructed image) to which in-loop filtering by a deep neural network filter or an existing in-loop filter is applied.

Encoding/decoding of deep neural network information

In operation 1830, encoding of the deep neural network information may be performed, and in operation 1930, decoding of the (encoded) deep neural network information may be performed.

In embodiments, encoding may mean entropy encoding. Decryption may mean entropy decoding.

Deep neural network information can be encoded and/or decoded within specific information within a bitstream. For example, specific information may include a sequence parameter set, a picture parameter set, and a slice header. Deep neural network information may be encoded and/or decoded between CTU syntax elements in a bitstream.

The deep neural network information may be information used to construct a deep neural network. The deep neural network information may be information representing a deep neural network. The deep neural network information may be information representing characteristics of the deep neural network.

The deep neural network information may include information for performing filtering using a filter and information for performing prediction using a predictive deep neural network.

The deep neural network information may include in-loop filter information and prediction deep neural network information.

The in-loop filter information includes: 1) filtering performance information for luma components, 2) filtering performance information for chroma components, 3) filter selection information for luma components, 4) filter selection information for chroma components, 5) specific target includes at least one of filtering performance information for , 6) filter selection information for a specific target, 7) deep neural network filter weight change information, 8) deep neural network filter output adjustment information, and 9) deep neural network filter weighted sum information. can do.

1) The filtering performance information on the luma component may indicate whether filtering is performed on the luma component.

2) The filtering performance information on the chroma component may indicate whether filtering is performed on the chroma component.

3) The filter selection information for the luma component may indicate a filter selected for the luma component.

4) The filter selection information for the chroma component may indicate a filter selected for the chroma component.

5) The filtering performance information for a specific target may indicate whether filtering is performed for the specific target.

In embodiments, a specific object may include a video, picture, slice, CTU, and block.

6) The filter selection information for a specific target may indicate a filter selected for the specific target.

7) The deep neural network filter weight change information may be information related to the weight of the deep neural network filter. The deep neural network filter weight change information may include weight information for the deep neural network filter and may include information indicating the number of weights.

8) The deep neural network filter output control information may include information used to control the output of the deep neural network filter. The deep neural network filter output control information may include information indicating whether to adjust the output of the deep neural network filter. The deep neural network filter output adjustment information may include a weight.

9) The deep neural network filter weighted sum information may include information used to derive a weighted sum of the output of the deep neural network filter and the result of the existing in-loop filter. The deep neural network filter weighted sum information may include weights of the weighted sum. The deep neural network filter weighted sum information may include information indicating a result used for a weighted sum among results of an existing in-loop filter.

The processing of at least one of filtering performance information on luma components, filter selection information on luma components, filtering performance information on chroma components, and filter selection information on chroma components may include filtering performance information on a specific target and filtering performance information on a specific target. It may be performed based on filter selection information.

For example, when the value of the filtering performance information for a specific target is the first value, filtering for the specific target may be performed. When the value of the filtering performance information for a specific target is the second value, filtering may not be performed for the specific target. In this case, filtering performance information for a specific target may be signaled for each specific target.

For example, when the value of the filtering performance information for a specific target is the first value, filter selection information for the specific target may be additionally signaled. Filtering for a specific target may be performed using the selected filter indicated by the filter selection information.

The predictive deep neural network information includes 1) predictive deep neural network performance information on luma components, 2) predictive deep neural network performance information on chroma components, 3) predictive deep neural network selection information on luma components, 4) predictive deep neural network performance information on chroma components. It may include one or more of selection information, 5) predictive deep neural network performance information for a specific target, 6) predictive deep neural network selection information for a specific target, and 7) predictive deep neural network weight change information.

1) The predictive deep neural network performance information on the luma component may indicate whether prediction using the predictive deep neural network is performed on the luma component.

2) The predictive deep neural network performance information on the chroma component may indicate whether prediction using the predictive deep neural network is performed on the chroma component.

3) The prediction deep neural network selection information for the luma component may indicate the prediction deep neural network selected for the luma component.

4) The prediction deep neural network selection information for the chroma component may indicate the prediction deep neural network selected for the chroma component.

5) The predictive deep neural network performance information for a specific object may indicate whether prediction using the predictive deep neural network is performed for the specific object.

In embodiments, a specific object may include a CTU and a block.

6) Prediction deep neural network selection information for a specific object may represent a predictive deep neural network selected for a specific object.

7) The prediction deep neural network weight change information may be information related to the weight of the prediction deep neural network. The prediction deep neural network weight change information may include weight information for the prediction deep neural network and may include information indicating the number of weights.

Processing of at least one of the predictive deep neural network performance information on the luma component, the predictive deep neural network selection information on the luma component, the predictive deep neural network performance information on the chroma component, and the predictive deep neural network selection information on the chroma component is performed for a specific target. It may be performed based on predictive deep neural network performance information and predictive deep neural network selection information for a specific target.

For example, when the value of the predictive deep neural network performance information for a specific target is the first value, prediction using the predictive deep neural network may be performed for the specific target. When the value of the predictive deep neural network performance information for a specific object is the second value, prediction using the predictive deep neural network may not be performed for the specific object. In this case, information about performing a predictive deep neural network for a specific target may be signaled.

For example, when the value of the predictive deep neural network performance information for a specific target is the first value, the predictive deep neural network selection information for the specific target may be additionally signaled. Prediction for a specific target may be performed using the selected predictive deep neural network indicated by the predictive deep neural network selection information.

For example, deep neural network information may include scaling information. A scaling operation may be applied to the output of the filter or predictive deep neural network to refine the output of the filter or predictive deep neural network. Fixed weights can be used to blend the output of the deep neural network filter and the output of the conventional in-loop filter. For example, the weight for a deep neural network filter may be 1, 0.75 or 0.5.

The above embodiments may be performed in the same method and/or a corresponding method in the encoding device 1600 and the decoding device 1700. Also, a combination of one or more of the above embodiments may be used in encoding and/or decoding an image.

The order in which the above embodiments are applied may be different in the encoding device 1600 and the decoding device 1700. Alternatively, the order in which the above embodiments are applied may be the same (at least partially) in the encoding device 1600 and the decoding device 1700.

The above embodiments may be performed for each of the luma signal and chroma signal. The above embodiments may perform the same with respect to the luma signal and the chroma signal.

The shape of a block to which the above embodiments are applied may have a square shape or a non-square shape.

Whether to apply and/or perform at least one of the above embodiments may be determined based on a condition for a block size. In other words, at least one of the above embodiments may be applied and/or performed when a condition for a block size is satisfied. The condition may include a minimum block size and a maximum block size. The block may be one of the blocks described above in the embodiments and the units described above in the embodiments. A block to which the minimum block size is applied and a block to which the maximum block size is applied may be different.

For example, when the size of a block is equal to or greater than the minimum size and/or when the size of a block is equal to or less than the maximum size, the above-described embodiment may be applied and/or performed. When the size of the block is greater than the minimum size and/or when the size of the block is less than or equal to the maximum size, the above-described embodiment may be applied and/or performed.

For example, the above-described embodiment may be applied only when the size of a block is a predefined block size. The predefined block size may be 2x2, 4x4, 8x8, 16x16, 32x32, 64x64 or 128x128. The predefined block size may be (2*SIZE _X )x(2*SIZE _Y ). SIZE _X can be one of 1 or more integers. SIZE _Y can be one of 1 or more integers.

For example, the above-described embodiment may be applied only when the size of a block is greater than or equal to the minimum block size. The above-described embodiment can be applied only when the size of a block is greater than the minimum block size. Block minimum size can be 2x2, 4x4, 8x8, 16x16, 32x32, 64x64 or 128x128. Alternatively, the minimum block size may be (2*SIZE _{MIN_X} )x(2*SIZE _{MIN_Y} ). SIZE _{MIN_X} can be one of 1 or more integers. SIZE _{MIN_Y} can be one of 1 or more integers.

For example, the above-described embodiment may be applied only when the size of a block is less than or equal to the maximum block size. The above-described embodiment can be applied only when the size of a block is smaller than the maximum block size. The maximum block size can be 2x2, 4x4, 8x8, 16x16, 32x32, 64x64 or 128x128. Alternatively, the maximum block size may be (2*SIZE _{MAX_X} )x(2*SIZE _{MAX_Y} ). SIZE _{MAX_X} can be one of 1 or more integers. SIZE _{MAX_Y} can be one of 1 or more integers.

For example, the above-described embodiment may be applied only when the size of a block is greater than or equal to the minimum block size and less than or equal to the maximum block size. The foregoing embodiment can be applied only when the block size is greater than the minimum block size and less than or equal to the maximum block size. The above-described embodiment can be applied only when the size of a block is greater than or equal to the minimum block size and smaller than the maximum block size. The above-described embodiment can be applied only when the block size is larger than the minimum block size and smaller than the maximum block size.

In the above-described embodiments, the size of a block may mean a horizontal size of a block or a vertical size of a block. The size of a block can mean both the horizontal size of the block and the vertical size of the block. Also, the size of a block may mean the area of the block. Each of the area, minimum block size, and maximum block size may be one of integers greater than or equal to 1. Also, the size of a block may mean a result (or value) of a known formula using the horizontal size and vertical size of a block or a result (or value) of a formula in an embodiment.

Also, in the above embodiments, the first embodiment may be applied to the first size, and the second embodiment may be applied to the second size.

The above embodiments may be applied according to a temporal layer. A separate identifier may be signaled to identify a temporal layer to which the above embodiments are applicable, and the above embodiments may be applied to a temporal layer specified by the identifier. The identifier herein may be defined as the lowest layer and/or the highest layer to which the above embodiment is applicable, or may be defined to indicate a specific layer to which the above embodiment is applied. In addition, a fixed temporal layer to which the above embodiment is applied may be defined.

For example, the above embodiments may be applied only when the temporal layer of the target image is the lowest layer. For example, the above embodiments may be applied only when the temporal layer identifier of the target image is 1 or more. For example, the above embodiments may be applied only when the temporal layer of the target image is the highest layer.

A slice type or tile group type to which the above embodiments are applied may be defined, and the above embodiments may be applied according to the corresponding slice type or tile group type.

In the above-described embodiments, in applying a specified process to a specified subject, a specified condition may be required, and if it is described that the specified process is processed under a specified decision, the specified coding parameter If it is described that whether or not a specified condition is satisfied based on a specific coding parameter or a specified decision is made based on a specified coding parameter, it can be interpreted that the specified coding parameter can be replaced with another coding parameter. That is to say, a coding parameter that affects a specified condition or a specified decision may be regarded as exemplary only, and it will be understood that a combination of one or more other coding parameters in addition to the specified coding parameter plays a role of the above specified coding parameter. can

In the foregoing embodiments, the methods are described on the basis of a flow chart as a series of steps or units, but the present invention is not limited to the order of steps, and some steps may occur in a different order or concurrently with other steps as described above. can In addition, those skilled in the art will understand that the steps shown in the flow chart are not exclusive, that other steps may be included, or that one or more steps of the flow chart may be deleted without affecting the scope of the present invention. You will understand.

The embodiments described above include examples of various aspects. While not all possible combinations for representing the various aspects may be described, those skilled in the art will recognize that other combinations than those explicitly described are possible. Accordingly, it is intended that the present invention cover all other substitutions, modifications and variations falling within the scope of the following claims.

Embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer readable recording medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in the art of computer software.

A computer-readable recording medium may contain information used in embodiments according to the present invention. For example, a computer-readable recording medium may include a bitstream, and the bitstream may include information described in embodiments according to the present invention.

A bitstream may include computer-executable code and/or programs. Computer-executable codes and/or programs may include information described in the embodiments and may include syntax elements described in the embodiments. In other words, the information and syntax elements described in the actual examples may be regarded as computer-executable code in a bitstream, and may be regarded as at least part of a computer-executable code and/or program represented in a bitstream. The recording medium may include a non-transitory computer-readable medium.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine language codes such as those produced by a compiler. The above hardware devices may be configured to act as one or more software modules to perform processing according to the present invention and vice versa.

In the above, the present invention has been described by specific details such as specific components and limited embodiments and drawings, but these are provided only to help a more general understanding of the present invention, and the present invention is not limited to the above embodiments. No, and those skilled in the art to which the present invention pertains may seek various modifications and variations from these descriptions.

Therefore, the spirit of the present invention should not be limited to the above-described embodiments and should not be determined, and not only the claims to be described later, but also all modifications equivalent or equivalent to these claims fall within the scope of the spirit of the present invention. would be said to belong.

Claims (20)

1. selecting a deep neural network;

   generating coded deep neural network information by encoding deep neural network information used to construct the deep neural network; and

   Generating a bitstream including the encoded deep neural network information

   An image encoding method comprising a.
2. According to claim 1,

   Performing learning in the deep neural network

   An image encoding method further comprising.
3. According to claim 2,

   The step of performing learning in the deep neural network,

   performing learning in the teacher network;

   conducting first learning in the student network; and

   performing a second learning in the student network;

   An image encoding method comprising a.
4. According to claim 3,

   Learning in the student network is performed using knowledge distillation using the teacher network.
5. According to claim 1,

   An image encoding method in which learning is performed in an in-loop filter based on the deep neural network.
6. According to claim 5,

   The input of the in-loop filter is an image in which compression deterioration has occurred,

   The output of the in-loop filter is an image from which compression degradation has been removed,

   An image encoding method in which learning is performed in the in-loop filter so that an error between an output of the in-loop filter and an original image is minimized.
7. According to claim 1,

   The deep neural network is a predictive deep neural network that generates a predicted image similar to an original image through prediction.
8. Receiving a bitstream including encoded deep neural network information;

   generating deep neural network information used to construct a deep neural network by decoding the encoded deep neural network information; and

   selecting the deep neural network

   Image decoding method comprising a.
9. According to claim 8,

   Performing learning in the deep neural network

   Image decoding method further comprising.
10. According to claim 9,

    The step of performing learning in the deep neural network,

    performing learning in the teacher network;

    conducting first learning in the student network; and

    performing a second learning in the student network;

    Image decoding method comprising a.
11. According to claim 10,

    Learning in the student network is performed using knowledge distillation using the teacher network.
12. According to claim 8,

    An image decoding method in which learning is performed in an in-loop filter based on the deep neural network.
13. According to claim 12,

    The input of the in-loop filter is an image in which compression deterioration has occurred,

    The output of the in-loop filter is an image from which compression degradation has been removed,

    An image decoding method in which learning is performed in the in-loop filter so that an error between an output of the in-loop filter and an original image is minimized.
14. According to claim 8,

    The deep neural network is a predictive deep neural network that generates a predicted image similar to an original image through prediction.
15. A computer-readable recording medium storing a bitstream for image decoding, wherein the bitstream comprises:

    Encoded Deep Neural Network Information

    including,

    Deep neural network information used to construct a deep neural network is generated by decoding the encoded deep neural network information;

    A computer-readable recording medium on which the deep neural network is selected.
16. According to claim 16,

    A computer-readable recording medium on which learning is performed in the deep neural network.
17. According to claim 16,

    Learning from the teacher network is performed,

    A first learning in the student network is performed;

    A computer-readable recording medium on which second learning is performed in the student network.
18. According to claim 17,

    Learning in the student network is performed using knowledge distillation using the teacher network.
19. According to claim 15,

    A computer-readable recording medium on which learning is performed in an in-loop filter based on the deep neural network.
20. According to claim 12,

    The input of the in-loop filter is an image in which compression deterioration has occurred,

    The output of the in-loop filter is an image from which compression degradation has been removed,

    A computer-readable recording medium in which learning is performed in the in-loop filter so that an error between an output of the in-loop filter and an original image is minimized.

Images