JP2021103519A - Method and system for normalizing smoothing feature of time space for behavior recognition - Google Patents

Abstract

To disclose a method and a system for normalizing smoothing features of a time space for behavior recognition.SOLUTION: A feature normalizing method includes a step of calculating a low-frequency component from an input feature, a step of calculating a high-frequency component by using a residual between the input feature and the low-frequency component, and a step of adding a noise to the low-frequency component.SELECTED DRAWING: Figure 3

Description

The following description relates to a feature normalization technique for action recognition.

Human behavior recognition technology is applied in many fields such as security-related fields such as intelligent video surveillance systems, intelligent robots capable of executing mutual exchange with humans, and intelligent home appliances.

For example, in Patent Document 1 (registration date: October 20, 2015), after extracting data necessary for behavior recognition using Kinect, the data is layered and features are learned to act from the video. The technology for recognizing is disclosed.

3D convolutional neural networks (3D Convolutional Neural Networks) are widely used in the field of behavior recognition. A 3D convolutional neural network is an extension of a 2D convolutional neural network (2D Conv Net) that has additional dimensions to handle spatiotemporal streams, and is a 2D kernel trained for large image recognition datasets. Inflate and utilize the knowledge learned in the image domain.

Korean Registered Patent No. 10-1563297

A simple and efficient normalization method can be provided to solve the problem of overfitting of a 3D convolutional neural network (3D Conv Net).

Random mean scaling (RMS) that normalizes the internal representation by randomly changing the magnitude of the low-frequency component of the feature can be applied.

A feature normalization method performed by a computer system, wherein the computer system includes at least one processor configured to execute a computer-readable instruction contained in memory, said feature normalization. The method is a step of obtaining a low frequency component from an input feature by the at least one processor, and a high frequency component is obtained by using the residual (residual) between the input feature and the low frequency component by the at least one processor. Provided is a feature normalization method comprising a step and a step of adding noise to the low frequency component by the at least one processor.

According to one aspect, in the step of obtaining the low frequency component, the low frequency component may be separated from the input feature by using a low-pass filter.

According to another aspect, the step of determining the low frequency component may utilize an average pooling or a Gaussian filter to separate the low frequency component from the input feature.

According to another aspect, the step of adding the noise may include a step of applying random scaling to the local average of the input features to add the noise.

According to another aspect, the step of adding the noise may include a step of randomly modulating the magnitude of the low frequency component by an operation of multiplying a scalar sampled with a given probability distribution.

According to another aspect, the random average scaling that adds the noise to the low frequency component may be applied within the residual branch of the network model.

According to another aspect, the random average scaling is at least one layer of the convolution layer, batch normalization layer, and non-linear activation layer of the network model. May be applied before.

According to another aspect, when the network model is a network with a basic block structure, the random average scaling is performed before all the batch normalization layers included in some stages of the network model, respectively. May be applied.

According to yet another aspect, if the network model is a network with a bottleneck block structure, the random average scaling is the last of the batch normalization layers included in some stages of the network model. May be applied before the batch normalization layer of.

Provided is a computer program recorded on a non-transitory computer-readable recording medium for causing the computer system to perform the feature normalization method.

Provided is a non-transitory computer-readable recording medium in which a program for causing a computer to execute the feature normalization method is recorded.

A computer system comprising at least one processor configured to execute a computer-readable instruction contained in memory, said at least one processor seeking a low frequency component from an input feature and with said input feature. Provided is a computer system characterized in that a high frequency component is obtained by utilizing a residual with the low frequency component and noise is added to the low frequency component.

According to an embodiment of the present invention, a simple and efficient normalization method can be provided to solve the problem of overfitting of a 3D convolutional neural network.

According to an embodiment of the present invention, the problem of overfitting of a 3D residual network (residual neural network) is solved by random mean scaling (RMS) that randomly changes the magnitude of the low frequency component of the feature to normalize the internal representation. It can be solved effectively.

According to an embodiment of the present invention, applying perturbation to a low frequency component can improve the accuracy and normalization effect of a 3D convolutional neural network as compared to applying it to an entire feature or high frequency component. ..

It is a block diagram for demonstrating an example of the internal structure of the computer system in one Embodiment of this invention. It is a graph which showed the accuracy of the 3D convolutional neural network with respect to the scaling factor in one Embodiment of this invention. It is a figure which showed the example of the random average scaling module for feature normalization in one Embodiment of this invention. It is a figure which showed the example of the random average scaling module for feature normalization in one Embodiment of this invention. It is a figure which showed the example of the network of the basic block structure to which the random average scaling module was added in one Embodiment of this invention. It is a figure which showed the example of the network of the bottleneck block structure to which the random average scaling module was added in one Embodiment of this invention. It is a figure which showed the experimental result which showed the normalization by the position of the random average scaling module in one Embodiment of this invention. It is a figure which showed the example of the component which can include the processor of the computer system in one Embodiment of this invention. It is a flowchart which showed the example of the feature normalization method which a computer system can perform in one Embodiment of this invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Embodiments of the present invention relate to feature normalization techniques for behavior recognition.

The embodiments, including those specifically disclosed herein, are simple and efficient by random mean scaling (RMS), which randomly changes the low frequency components of the features to normalize the internal representation. A normalization method can be provided, which can improve the accuracy and normalization performance of behavior recognition.

FIG. 1 is a block diagram showing an example of a computer system according to an embodiment of the present invention. For example, the feature normalization system according to the embodiment of the present invention may be realized by the computer system 100 shown in FIG.

As shown in FIG. 1, the computer system 100 includes a memory 110, a processor 120, a communication interface 130, and an input / output interface 140 as components for executing the feature normalization method according to the embodiment of the present invention. Is fine.

The memory 110 is a computer-readable recording medium and may include a permanent large-capacity recording device such as a RAM (random access memory), a ROM (read only memory), and a disk drive. Here, a permanent large-capacity recording device such as a ROM or a disk drive may be included in the computer system 100 as another permanent recording device that is separated from the memory 110. Further, the memory 110 may record an operating system and at least one program code. Such software components may be loaded into memory 110 from a computer-readable recording medium separate from memory 110. Such other computer readable recording media may include computer readable recording media such as floppy® drives, discs, tapes, DVD / CD-ROM drives, memory cards and the like. In other embodiments, software components may be loaded into memory 110 through a communication interface 130 that is not a computer-readable recording medium. For example, software components may be loaded into memory 110 of computer system 100 based on a computer program installed by a file received over network 160.

Processor 120 may be configured to process instructions in a computer program by performing basic arithmetic, logic, and input / output operations. Instructions may be provided to processor 120 by memory 110 or communication interface 130. For example, the processor 120 may be configured to execute instructions received according to program code recorded in a recording device such as memory 110.

The communication interface 130 may provide a function for the computer system 100 to communicate with other devices via the network 160. As an example, requests, instructions, data, files, etc. generated by the processor 120 of the computer system 100 according to a program code recorded in a recording device such as a memory 110 can be transmitted via the network 160 under the control of the communication interface 130. May be transmitted to the device of. On the contrary, signals, commands, data, files and the like from other devices may be received by the computer system 100 via the communication interface 130 of the computer system 100 via the network 160. Signals, instructions, data and the like received through the communication interface 130 may be transmitted to the processor 120 and the memory 110, and the files and the like may be further included in a recording medium (the above-mentioned permanent recording device) that the computer system 100 can include. May be recorded.

The communication method is not limited, and not only the communication method using the communication network (for example, mobile communication network, wired Internet, wireless Internet, broadcasting network) that can be included in the network 160, but also the short distance between devices. Wireless communication may be included. For example, the network 160 includes a PAN (personal area network), a LAN (local area network), a CAN (campus area network), a MAN (metropolitan area network), a WAN (wide network), etc. It may include any one or more of the networks. Further, network 160 may include, but is limited to, any one or more of network topologies, including bus networks, star networks, ring networks, mesh networks, star-bus networks, tree or hierarchical networks, and the like. Will not be done.

The input / output interface 140 may be a means for interfacing with the input / output device 150. For example, an input device may include a device such as a microphone, keyboard, camera, or mouse, and an output device may include a device such as a display or speaker. As another example, the input / output interface 140 may be a means for an interface with a device such as a touch screen in which functions for input and output are integrated into one. The input / output device 150 may be composed of a computer system 100 and one device.

Also, in other embodiments, the computer system 100 may include fewer or more components than the components of FIG. However, it is not necessary to clearly illustrate most of the prior art components. For example, the computer system 100 may be implemented to include at least a portion of the input / output devices 150 described above, and may further include other components such as transceivers, cameras, various sensors, databases, and the like. It may be.

In the field of video behavior recognition, deep neural networks (DNNs) often require a 3D convolution filter, and overfitting often occurs due to a large number of parameters.

One of the problems facing DNN, overfitting, is the field of video behavior recognition, where 3D convolutional neural networks (3D Conv Net) are the preferred approach for encoding spatio-temporal representation. Is especially fatal.

Despite the ability of 3D convolutional neural networks to process video streams, many parameters often face excessive problems.

Normalization in the input space and the feature space is a well-known approach to alleviate the problem of overfitting, but previous studies have motivated where the effect on the feature begins. The direction of slippage (direction to perturb) is overlooked.

Since we assume that such a direction is a very important factor for normalization, it is necessary to analyze what kind of information is important for behavior recognition tasks.

FIG. 2 is a diagram showing changes in the accuracy of a 3D convolutional neural network with respect to various scaling factors that modulate the magnitude of the low and high frequency components of a feature. According to FIG. 2, it can be seen that the behavior recognition performance is more sensitive to the high frequency component than to the low frequency.

Random mean scaling (RMS) is applied as a normalization method in embodiments of the present invention, based on the fact that selective perturbation on features is an effective way to normalize the network.

The Random Mean Scaling (RMS) method according to an embodiment of the present invention multiplies a random scalar by a spatiotemporal smoothing feature to selectively add perturbations. To separate low frequency information, a 3D averaging filter (3D averaging pooling operation in most deep learning), which is the simplest low-pass filter in image processing, may be used. Like other normalization methods, the Random Mean Scaling (RMS) method is required only during training and requires no additional work during inference.

First, behavior recognition and normalization will be described.

The behavior recognition 3D convolutional neural network is extended by a 2D convolutional neural network (2D ConvNet) having an additional dimension for processing a spatiotemporal stream, and is widely used in the field of behavior recognition. A 3D convolutional neural network may inflate a 2D kernel trained on a large image recognition dataset to leverage the knowledge learned in the image domain.

However, 3D convolutional neural networks have many parameters as their disadvantages. In order to overcome such a problem, the 3D kernel may be decomposed into a stepped structure consisting of a 2D kernel and a 1D kernel. As an example, the 3D filter may be decomposed into 2D filters that can be applied simultaneously by HW, TH, TW, and only the 3D convolution filter may be used in the next step.

On the other hand, some studies have proposed a multi-step model, which uses information individually according to frequency. For example, two stream models may be utilized that consist of a slow branch for static spatial features and a fast branch for dynamic motion features. Octave convolution may also be used to process multiple frequency signals in a single stream model.

Capturing global dependency is another approach to improving behavioral cognitive models. For example, a non-local module may be added as a method for reducing the number of operations of a 3D convolutional neural network.

Normalization Normalization is effective in eliminating model overfitting, but it has been less active in the video domain than in the image domain. Image domains primarily use normalization techniques such as data augmentation, weight decay, dropout, label smoothing, and batch normalization. ing.

Recent studies have made it possible to extend data to the input data space by methods such as random interpolation, interpolating two images or transplating an image patch onto another image.

Also, internal representation is a separate subject from normalization in recent studies. Shake-Shake normalization techniques normalize multi-branch ResNets by adding randomly scaled branches to forward and backward operations that are not applicable to 2-branch ResNets. Also, the stochastic depth (in other words, RandomDrop) technique randomly switches between a main branch and a residual branch connected to it. Shakedrop, which combines these two technologies, adopts a random drop conversion mechanism as Shake-Shake, and is compatible with 2-branched ResNet.

Hereinafter, the random scaling method for the spatiotemporal smoothing feature in the embodiment of the present invention will be described.

Random mean scaling (RMS)
Random scaling is a simple normalization method that can be applied to any hierarchy of convolutional neural networks.

Random scaling is a method of randomly modulating the size of a feature by multiplying it by a scalar α sampled with a given probability distribution (eg, Gaussian).

In this embodiment, random scaling is not applied directly to the features, but to the local mean of the features (random average scaling). In particular, the random average scaling method separates the inputs into frequency features (high frequency and low frequency components) to reduce overfitting and randomly adds noise to the low frequency components.

Generally, in an image, the high frequency component contains edge information, which can be said to be important information for classification. Therefore, the normalization effect can be improved by randomly scaling only the low-frequency components without changing the high-frequency components that are important for classification.

A low-pass filter may be used to separate the low frequency components from the input, for example, an average pooling or box filter, a Gaussian filter, or the like may be used.

The local average may be calculated as in formula (1).

Here, x denotes the input features, _{W i} refers to 3D local window of the current index i surrounding the (local window).

^{The input feature x is separated into an average x −} and a residual r as in equation (2).

The modulation output y by random average scaling may be defined as in equation (3).

Perturbations may only be applied during training. If the average of the probability distributions of α is 1, then y = x during inference.

The random average scaling method described above is applicable to any level of hierarchy such as convolution, batch normalization (BN), nonlinear activation (nonlinear activation), and the like.

Random average scaling positions may be determined to improve network performance. Also, applying random scaling to the local mean can improve performance over applying random scaling to the residuals or the entire input. The local average is described as the low frequency component of the input, while the residual indicates the remaining high frequency component.

As explained with reference to FIG. 2, since high frequency modulation significantly reduces performance, random average scaling is presumed to generate a model to make better use of high frequency components compared to low frequencies.

FIG. 3 is a diagram showing a random average scaling module 300 according to an embodiment of the present invention. Random average scaling may be implemented by a network module that performs some basic operations, as shown in FIG.

は要素ごとの和（ｅｌｅｍｅｎｔ−ｗｉｓｅ ｓｕｍ）の演算を示し、

は要素ごとの積（ｅｌｅｍｅｎｔ−ｗｉｓｅ ｍｕｌｔｉｐｌｉｃａｔｉｏｎ）の演算を示す。 In FIG. 3, x is the input, x ⁻ is the input mean, r is the residual, and y is the output.

Indicates the sum (element-wise sum) operation for each element.

Indicates an operation of the product (elent-wise multiplication) for each element.

In FIG. 3, for convenience of explanation, a random average scaling module 300 having a structure in which the ^{input average x − and the residual r are separated is shown.} The local average of formula (1) can be determined by the 3D average pooling provided by most deep learning frameworks, and formula (3) is shown as shown in FIG.

For actual realization, the mathematical formula (3) may be modified to a simpler form while using the mathematical formula (2) (mathematical formula (4)).

Here, α'= α-1.

FIG. 4 shows a simplified form for realizing the random average scaling module, and shows a block diagram showing the random average scaling module 300 corresponding to the mathematical formula (4).

As mentioned above, the random average scaling module 300 requires only simple operations such as multiplication of average pooling and scalars, so it has no parameters and requires only a small amount of additional operations during training. No additional operations are required during inference.

An example of a network whose effect is improved by random average scaling is the 3D ResNet series. For example, SlowOnly, CSN (channel-separated parameter network), etc. correspond to this, but in SlowOnly, 3D convolution is used only in res4 and res5 stages in a form in which general 2D ResNet is firmly set in 3D. It can be said that the CSN is a light-weight (fewer parameters) 3D ResNet because there is no dimension reduction due to the axis.

The random average scaling module 300 can be applied to any position in the residual branch and may be located, for example, in front of the convolution layer or the ReLu layer.

As an example, FIG. 5 shows an example of a network having a basic block structure. In the case of the basic block structure, a random average scaling module 300 may be added in front of all the BN (batch normalization) layers included in the res4 and res5 stages, respectively.

As another example, FIG. 6 shows an example of a network having a bottleneck block structure. In the case of the bottleneck block, the random average scaling module 300 may be added before the last BN (batch normalization) layer of the BN (batch normalization) layers included in the res4 and res5 stages.

The random average scaling module 300 can be applied to any level, such as before each convolution layer, before each BN (batch regularization) layer, before each ReLU layer, and so on. The random average scaling module 300 before the last ReLU layer may be located before the sum of the main branch and the residual branch.

The experimental results of the normalization effect of each position of the random average scaling module 300 are as shown in FIG.

In FIG. 7, when the result of adding the random average scaling module 300 to all possible positions with respect to SlowOnly-34 is examined in detail, the difference in the normalization effect depending on the position where the random average scaling module 300 is added is not large. It can be seen that, among the cases where a single random average scaling module 300 is used, the random average scaling module 300 before the first BN shows the highest accuracy.

Furthermore, in the case of the bottleneck block structure SlowOnly-50, the random average scaling module 300 improved the performance of the network in all cases, and in particular, the random average scaling module 300 before the last BN showed the highest accuracy. .. The random average scaling module 300 before the first BN is also an efficient choice because there is no big difference in performance, but for computational efficiency, the bottleneck block structure uses multiple random average scaling modules 300. You may choose not to.

Therefore, it was found that the random average scaling module 300 exhibits higher performance than the past models not only in the basic block structure but also in the bottleneck block structure.

FIG. 8 is a diagram showing an example of components that can be included in the processor of the computer system according to the embodiment of the present invention, and FIG. 9 is a diagram showing the execution by the computer system according to the embodiment of the present invention. It is a flowchart which showed the example of the feature normalization method which can be done.

As shown in FIG. 8, the processor 120 may include a frequency separation unit 801 and a normalization unit 802. Such components of the processor 120 may be representations of different functions that are executed by the processor 120 according to control instructions provided by at least one program code. For example, the frequency separator 801 may be used as a functional representation in which the processor 120 operates to control the computer system 100 to separate the inputs into frequency characteristics.

Processor 120 and the components of processor 120 may perform steps 910-920 included in the feature normalization method of FIG. For example, the processor 120 and the components of the processor 120 may be implemented to execute instructions by the operating system code included in the memory 110 and at least one program code described above. Here, at least one program code may correspond to the code of the program implemented to handle the feature normalization method.

The feature normalization method may not occur in the order shown in the figure, some of the steps of the method may be omitted, or additional steps may be included.

The processor 120 may load the program code recorded in the program file for the feature normalization method into the memory 110. For example, the program file for the feature normalization method may be recorded in a persistent recording device that is distinct from memory 110, and the processor 120 may be recorded from the program file recorded in the persistent recording device via the bus. The computer system 100 may be controlled so that the program code is loaded into the memory 110. At this time, the processor 120 and the frequency separation unit 801 and the normalization unit 802 included in the processor 120 each execute the instruction of the corresponding part of the program code loaded in the memory 110 to execute the following steps 910 to 920. It may be a different functional representation of the processors 120 for the purpose. For the execution of steps 910-920, the processor 120 and the components of the processor 120 may directly process the operations by the control instructions or control the computer system 100.

Processor 120 may include the random average scaling module 300 described with reference to FIGS. 3 and 4.

The feature normalization method according to the present invention may include the following two steps.

At step 910, the frequency separator 801 may separate the input features into frequency characteristics. The frequency separation unit 801 may obtain a low frequency component from the input feature by using a low-pass filter, and obtain a high frequency component by using the residual between the input feature and the low frequency component. As an example, the frequency separation unit 801 may separate the feature map into a high frequency component (residual) and a low frequency component (local average) by using average pooling.

At step 920, the normalization unit 802 may normalize the input features by randomly adding noise to the low frequency components separated from the input features. The normalization unit 802 applies random average scaling (RMS) to solve the problem of overfitting of the 3D residual network, so that the high frequency component and the low frequency component separated from the input features are high frequency components. May be maintained and low frequency components may be scaled randomly. For example, the normalization unit 802 may randomly change the low-frequency portion of the feature by utilizing a uniform distribution, a normal distribution, or the like.

Therefore, the processor 120 can easily and efficiently solve the problem of overfitting by separating the low frequency components in the feature map and randomly changing the corresponding low frequency components to normalize the internal representation. Can be done. Finally, normalization for smoothed features can lead to performance gains by selectively explicitly handling low and high frequency components.

Thus, according to an embodiment of the present invention, the problem of overfitting of a 3D residual network is solved by random mean scaling (RMS), which randomly changes the magnitude of the low frequency components of a feature to normalize the internal representation. It can be solved effectively. Further, according to the embodiment of the present invention, by applying the perturbation to the low frequency component, the accuracy and normalization effect of the 3D convolutional neural network can be improved as compared with the application to the entire feature or the high frequency component.

The devices described above may be implemented by hardware components, software components, and / or combinations of hardware components and software components. For example, the devices and components described in the embodiments include a processor, a controller, an ALU (arithmetic logic unit), a digital signal processor, a microcomputer, an FPGA (field programgate array), a PLU (programmable log unit), a microprocessor, and the like. Alternatively, it may be implemented using one or more general purpose computers or special purpose computers, such as various devices capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the OS. The processing device may also respond to the execution of the software, access the data, and record, manipulate, process, and generate the data. For convenience of understanding, one processor may be described as being used, but one of ordinary skill in the art may appreciate that the processor may include multiple processing elements and / or multiple types of processing elements. You can understand. For example, the processing device may include multiple processors or one processor and one controller. Other processing configurations, such as parallel processors, are also possible.

The software may include computer programs, code, instructions, or a combination of one or more of these, configuring the processing equipment to operate at will, or instructing the processing equipment independently or collectively. You may do it. The software and / or data is embodied in any type of machine, component, physical device, computer recording medium or device to be interpreted based on the processing device or to provide instructions or data to the processing device. Good. The software is distributed on a networked computer system and may be recorded or executed in a distributed state. The software and data may be recorded on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed by various computer means and recorded on a computer-readable medium. Here, the medium may be a continuous recording of a computer-executable program or a temporary recording for execution or download. Further, the medium may be various recording means or storage means in the form of a combination of a single piece of hardware or a plurality of pieces of hardware, and is not limited to a medium directly connected to a certain computer system, but is distributed on a network. It may exist. Examples of media include hard disks, floppy (registered trademark) disks, magnetic media such as magnetic tapes, optical media such as CD-ROMs and DVDs, optical magnetic media such as floptic discs, and the like. And ROM, RAM, flash memory, etc., and may be configured to record program instructions. Other examples of media include recording media or storage media managed by application stores that distribute applications, sites that supply or distribute various other software, servers, and the like.

As described above, the embodiments have been described based on the limited embodiments and drawings, but those skilled in the art will be able to make various modifications and modifications from the above description. For example, the techniques described may be performed in a different order than the methods described, and / or components such as the systems, structures, devices, circuits described may be in a form different from the methods described. Appropriate results can be achieved even if they are combined or combined, or confronted or replaced by other components or equivalents.

Therefore, even if the embodiments are different, they belong to the attached claims as long as they are equal to the claims.

120: Processor 801: Frequency separation unit 802: Normalization unit

Claims (20)

A feature normalization method performed by a computer system
The computer system includes at least one processor configured to execute computer-readable instructions contained in memory.
The feature normalization method is
The step of obtaining a low frequency component from an input feature by the at least one processor.
A step of obtaining a high frequency component by utilizing the residual of the input feature and the low frequency component by the at least one processor.
A feature normalization method comprising adding noise to the low frequency components by the at least one processor. The step of obtaining the low frequency component is
The feature normalization method according to claim 1, wherein the low-frequency component is separated from the input feature by using a low-pass filter. The step of obtaining the low frequency component is
The feature normalization method according to claim 1, wherein the low frequency component is separated from the input feature by using an average pooling or a Gaussian filter. The stage of adding the noise is
The feature normalization method according to claim 1, wherein the noise is added by applying random scaling to the local average of the input features. The stage of adding the noise is
The feature normalization method according to claim 1, further comprising a step of randomly modulating the magnitude of the low frequency component by an operation of multiplying a scalar sampled with a given probability distribution. The feature normalization method according to claim 1, wherein the random average scaling that adds the noise to the low frequency component is applied within the residual branch of the network model. The feature normalization according to claim 6, wherein the random average scaling is applied before at least one layer of the convolution layer, the batch normalization layer, and the nonlinear activation layer of the network model. Method. If the network model is a network with a basic block structure, the random average scaling is applied before all the batch normalization layers included in some stages of the network model, respectively. Item 6. The feature normalization method according to Item 6. If the network model is a network with a bottleneck block structure, the random average scaling is applied before the last batch normalization layer of the batch normalization layers included in some stages of the network model. The feature normalization method according to claim 6, wherein the feature normalization method is characterized in that. A computer program that causes the computer system to execute the feature normalization method according to any one of claims 1 to 9. A non-transitory computer-readable recording medium in which a program for causing a computer to execute the feature normalization method according to any one of claims 1 to 9 is recorded. It ’s a computer system,
Contains at least one processor configured to execute computer-readable instructions contained in memory.
The at least one processor
Find the low frequency component from the input features
The high frequency component is obtained by using the residual between the input feature and the low frequency component.
A computer system characterized by adding noise to the low frequency components. The at least one processor
The computer system according to claim 12, wherein a low-pass filter is used to separate the low-frequency component from the input feature. The at least one processor
12. The computer system of claim 12, characterized in that the low frequency components are separated from the input features using an average pooling or Gaussian filter. The at least one processor
12. The computer system of claim 12, wherein random scaling is applied to the local average of the input features to add the noise. The at least one processor
The computer system according to claim 12, wherein the magnitude of the low frequency component is randomly modulated by an operation of multiplying a scalar sampled with a given probability distribution. The at least one processor
Includes a random average scaling module that adds the noise to the low frequency components.
The computer system according to claim 12, wherein the random average scaling module is located within a residual branch of the network model. The computer system according to claim 17, wherein the random average scaling module is located in front of at least one layer of a convolution layer, a batch normalization layer, and a nonlinear activation layer. When the network model is a network with a basic block structure, the random average scaling module is respectively located in front of all batch normalization layers included in some stages of the network model. Item 17. The computer system according to item 17. When the network model is a network with a bottleneck block structure, the random average scaling module is located before the last batch normalization layer among the batch normalization layers included in some stages of the network model. The computer system according to claim 17, wherein the computer system is characterized in that.

Images