CN117011299A - Reference point cloud quality assessment method and system integrating graph resampling and gradient characteristics - Google Patents

Abstract

The application discloses a reference point cloud quality assessment method and a system for merging graph resampling and gradient characteristics, which relate to the field of image processing, and comprise the following steps: extracting key points from the input reference point cloud by using a graph-based key point resampling method; dividing a local neighborhood group of a reference point cloud and a distortion point cloud by taking the key point as a center, and clustering the points in each local area according to Euclidean distance between the key point and other points in a coordinate space; respectively extracting three-dimensional gradient amplitude characteristics and three-dimensional gradient pattern characteristics of the reference point cloud and the distortion point cloud, calculating three-dimensional gradient amplitude similarity based on the three-dimensional gradient amplitude characteristics, calculating three-dimensional gradient pattern similarity based on the three-dimensional gradient pattern characteristics, and calculating joint three-dimensional gradient characteristic similarity; and based on the similarity of the combined three-dimensional gradient characteristics, carrying out weighted pooling by using the response intensity values to obtain the objective quality score of the distorted point cloud. The application has higher identification accuracy, sensitivity and robustness.

Description

Reference point cloud quality assessment method and system integrating graph resampling and gradient features

Technical field

The invention relates to the field of image processing, and in particular to a reference point cloud quality assessment method and system that integrates graph resampling and gradient features.

Background technique

With the continuous development of imaging technology, the demand for three-dimensional data in the field of computer vision is increasing. Due to its flexible representation, 3D point cloud shows wide application prospects in modern communication systems and has become one of the commonly used three-dimensional data formats in immersive media applications. Three-dimensional data contains depth information, which brings a more intuitive sense of space and three-dimensionality in visual expression, which is in line with the human visual system's perception and understanding of objects during observation.

However, point clouds are affected by noise and distortion during data processing, which causes distortion and reduces the visual quality of the displayed content for the human eye, thus affecting the end user's satisfaction with the visual experience. Therefore, designing a point cloud quality assessment method that takes into account the special structural characteristics of point clouds and conforms to human visual characteristics has important application value for the steady development of point cloud data collection, processing and application. Point cloud quality assessment has a greater significance currently. It has high theoretical research significance and practical application value.

Contents of the invention

The main purpose of the present invention is to overcome the defects in the existing technology and propose a reference point cloud quality assessment method and system that integrates graph resampling and gradient features, effectively describing the gradient direction features and response intensity caused by point cloud distortion. The changes are in line with the human eye's subjective perception of distorted point clouds, and have high recognition accuracy, sensitivity and robustness.

The present invention adopts the following technical solutions:

On the one hand, a reference point cloud quality assessment method that combines graph resampling and gradient features includes:

S101, use the graph-based key point resampling method to extract key points from the input reference point cloud;

S102, divide the local neighborhood group of the reference point cloud and the distortion point cloud with the key point as the center, and cluster the points in each local area according to the Euclidean distance between the key point and other points in the coordinate space;

S103, respectively extract the three-dimensional gradient amplitude features and the three-dimensional gradient direction map features of the reference point cloud and the distorted point cloud in the local area, and calculate the three-dimensional gradient amplitude similarity of the reference point cloud and the distorted point cloud based on the three-dimensional gradient amplitude features. The three-dimensional gradient pattern feature calculates the three-dimensional gradient pattern similarity between the reference point cloud and the distorted point cloud, and calculates the joint three-dimensional gradient feature similarity based on the three-dimensional gradient amplitude similarity and the three-dimensional gradient pattern similarity;

S104, based on the joint three-dimensional gradient feature similarity, use the response intensity value to perform weighted pooling to obtain the objective quality score of the distorted point cloud.

Preferably, the S101 specifically includes:

Reference point cloud for input , use high-pass image filtering method to // Perform resampling to extract key points; specifically use the graph transfer operator/> To describe the relative position relationship between nodes,/> Represents graph transfer operator/> Dimensions; use adjacency matrix/> and transfer matrix/> to express/> , that is/> , every linear shift invariant graph filter is/> polynomial function:

in, Represents a linear shift invariant graph filter;/> Represents the linear shift invariant graph filter./> factors,/> Represents the graph transfer operator /> Power;/> Indicates the first/> coefficient;/> Indicates the length of the graphics filter;/> represents the degree matrix;/> Express/> The inverse matrix;/> represents the identity matrix;

Implemented using Harr-like filters The high-pass filtering is as follows:

in, Represents the high-pass filtering of Har-like filter;/> Express/> of/> feature vector;/> Express/> eigenvalues; in the graph vertex domain, the Click/> The frequency response formula of is:

in, Indicates the first/> Click/> The frequency response formula;/> Represents the total number of points in the spatial frequency domain;/> Indicates the first/> Click/> adjacent points;

For input reference point cloud Extract key points to generate skeleton point cloud/> .

Preferably, the S102 specifically includes:

The resampled key point skeleton point cloud Each point in is used as a reference point cloud/> and distorted point cloud/> The local area center of /> No./> in key points/> , use/> and/> Euclidean distance pairs of corresponding geometric parts in/> Neighbors of points are clustered, and the formula is as follows:

in, Represented as key points/> is the center of the ball,/> is the neighborhood group of the reference point cloud with radius;/> Represented as key points/> is the center of the ball,/> is the neighborhood group of the distorted point cloud with radius;/> Represents the set of all points in the key point skeleton of the reference point cloud;/> Represents the set of all points in the key point skeleton of the distorted point cloud;/> represents the square of the second norm.

Preferably, the S103 specifically includes:

According to the reference point cloud The three-dimensional gradient amplitude characteristics/> with three-dimensional gradient pattern/> , and distorted point cloud/> The three-dimensional gradient amplitude characteristics/> with gradient pattern/> , calculate the three-dimensional gradient amplitude similarity of the reference point cloud and the distorted point cloud/> And three-dimensional gradient direction similarity/> ,Calculated as follows:

in, and/> is a constant used to ensure numerical stability;

For points in the local area of the reference point cloud , three-dimensional gradient amplitude characteristics/> Calculated as follows:

in, ,/> ,/> Respectively represent the gradient amplitude differences of points in the local neighborhood group of the reference point cloud in the three directions of x, y, and z;/> Represents the gradient calculation method on the xyz axis of points within the local neighborhood group of the reference point cloud;

The three-dimensional gradient direction is introduced to describe the change information in the point cloud space, with key points is the center of the sphere,/> Constructs a local area for the radius, for points falling within the local area/> , the formula for calculating the spherical coordinate value of this point is as follows:

in, Indicates point/> The polar angle of /> Indicates point/> Azimuth angle, perform direction clustering based on the spherical coordinate values of all points in the local area; set the eight directions of the three-dimensional gradient. The closer the spherical coordinate values of the points in the local area are to the direction, the clustering will be in that direction; the three-dimensional gradient The eight directions are calculated based on the polar angle and azimuth angle, and are equally spaced in space; the specific value calculation formula is:

in ;/> Indicates polar angle/> direction,/> represents the azimuth angle direction, the two directions combine with each other to form a three-dimensional gradient eight directions;

3D gradient pattern Calculated as follows:

in, for point/> Convolution kernel in three-dimensional eight directions. The direction of the convolution kernel depends on the direction of the three-dimensional gradient in which the point is clustered;/> Represents the convolution operation;

For points in the local area of the distorted point cloud , three-dimensional gradient amplitude characteristics/> Calculated as follows:

in, ,/> ,/> Respectively represent the gradient amplitude differences of points in the local neighborhood group of the distorted point cloud in the three directions of x, y, and z;/> Represents the gradient calculation method on the xyz axis of points within the local neighborhood group of the distorted point cloud;

The three-dimensional gradient direction is introduced to describe the change information in the point cloud space, with key points is the center of the sphere,/> Constructs a local area for the radius, for points falling within the local area/> , the formula for calculating the spherical coordinate value of this point is as follows:

in, Indicates point/> The polar angle of /> Indicates point/> Azimuth angle, perform direction clustering based on the spherical coordinate values of all points in the local area; set the eight directions of the three-dimensional gradient. The closer the spherical coordinate values of the points in the local area are to the direction, the clustering will be in that direction; the three-dimensional gradient The eight directions are calculated based on the polar angle and azimuth angle, and are equally spaced in space; the specific value calculation formula is:

in ;/> Indicates polar angle/> direction,/> represents the azimuth angle direction, the two directions combine with each other to form a three-dimensional gradient eight directions;

3D gradient pattern Calculated as follows:

in, for point/> Convolution kernel in eight three-dimensional directions. The direction of the convolution kernel depends on the direction of the three-dimensional gradient in which the point is clustered;

The joint three-dimensional gradient feature similarity is calculated based on the three-dimensional gradient amplitude similarity and the three-dimensional gradient direction map similarity, as follows:

Among them, parameters is used to adjust/> and/> A positive number of relative importance between.

Preferably, the S104 is as follows:

in, is the response intensity value of points in the local area of the reference point cloud;/> The response intensity value of points in the local area of the distorted point cloud;/> Indicates the number of local areas;/> Represents the local range of the reference point cloud and distortion point cloud, that is, the three-dimensional point xyz falls within this local range;/> Represents the objective quality score of the distorted point cloud.

Preferably, the calculation method of the response intensity value is a K nearest neighbor search algorithm.

On the other hand, a reference point cloud quality assessment system that integrates graph resampling and gradient features includes:

The key point extraction module is used to extract key points from the input reference point cloud using the graph-based key point resampling method;

The local area point clustering module is used to divide the local neighborhood group of the reference point cloud and the distortion point cloud with the key point as the center, and cluster each local area according to the Euclidean distance between the key point and other points in the coordinate space. Points within the region;

The gradient feature similarity calculation module is used to extract the three-dimensional gradient amplitude features and the three-dimensional gradient pattern features of the reference point cloud and the distorted point cloud in the local area, and calculate the three-dimensional gradient feature of the reference point cloud and the distorted point cloud based on the three-dimensional gradient amplitude feature. Gradient amplitude similarity, calculate the three-dimensional gradient pattern similarity between the reference point cloud and the distorted point cloud based on the three-dimensional gradient pattern feature, and calculate the joint three-dimensional gradient feature similarity based on the three-dimensional gradient amplitude similarity and the three-dimensional gradient pattern similarity. ;

The quality assessment module is used to obtain an objective quality score of the distorted point cloud based on the joint three-dimensional gradient feature similarity and weighted pooling using response intensity values.

Compared with the prior art, the beneficial effects of the present invention are as follows:

(1) The present invention extracts the geometric information of the reference point cloud through a graph-based resampling method to generate a key point skeleton, builds a local area parameterized model with the key point as the center of the sphere for the reference point cloud and the distortion point cloud, and measures the gradient respectively. The similarity of the amplitude and gradient direction is finally used to perform weighted pooling on the similarity of the two features using the weighted merger calculation method of the response intensity value, and the point cloud quality evaluation score is calculated, which effectively describes the point cloud distortion caused by The gradient direction characteristics and response intensity changes are in line with the human eye's subjective perception of distorted point clouds, with high recognition accuracy, sensitivity and robustness, and good point cloud quality assessment performance;

(2) The present invention extracts the geometric information of the reference point cloud through a graph-based resampling method to generate a key point skeleton, which solves the problem of difficult feature mapping of reference and distorted point clouds in quality assessment. The sampled key points are not only in line with human perception, and significantly reduces computational complexity.

Description of the drawings

Figure 1 is a flow chart of a reference point cloud quality assessment method for fusion graph resampling and gradient features according to an embodiment of the present invention;

Figure 2 is a structural flow chart of a reference point cloud quality assessment method for fusion graph resampling and gradient features according to an embodiment of the present invention;

Figure 3 is a structural block diagram of a reference point cloud quality assessment system that fuses graph resampling and gradient features according to an embodiment of the present invention.

Detailed ways

The present invention will be further described below in conjunction with specific embodiments. It should be understood that these examples are only used to illustrate the invention and are not intended to limit the scope of the invention. In addition, it should be understood that after reading the teachings of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

Referring to Figures 1 and 2, a reference point cloud quality assessment method that fuses graph resampling and gradient features of the present invention is shown. The specific steps are as follows:

S101, use the graph-based key point resampling method to extract key points from the input reference point cloud.

Specifically, receive the input reference point cloud , using a high-pass image filtering method to // Resampling is performed to extract key points. Use graph transfer operator/> To describe the relative position relationship between nodes,/> Represents graph transfer operator/> Dimensions;/> You can use adjacency matrix/> and transfer matrix/> to express, that is/> . Every linear shift invariant graph filter is/> polynomial function:

in, Represents a linear shift invariant graph filter;/> Represents the graph transfer operator /> Power;/> Indicates the first/> coefficient;/> Indicates the length of the graphics filter;/> represents the degree matrix;/> Express/> The inverse matrix;/> represents the identity matrix. In this embodiment, a Harr-like filter is used to implement/> High pass filtering:

in, Represents the high-pass filtering of Har-like filter;/> Express/> of/> feature vector;/> Express/> eigenvalues; in the graph vertex domain, the Click/> The frequency response formula (referring to a point in the frequency domain of the point cloud space) is:

in, Indicates the first/> Click/> The frequency response formula;/> Represents the total number of points in the spatial frequency domain;/> Indicates the first/> Click/> adjacent points;

is used to reference point clouds in the spatial frequency domain/> is sorted for sampling, which reflects the How much information about the point can be collected from adjacent points? The more information a point has, the more likely it is to be selected as a key point. Finally take the filter length/> and the number of effective key points after resampling/> , to achieve a balance between complexity and efficiency.

S102, divide the local neighborhood group of the reference point cloud and the distortion point cloud with the key point as the center, and cluster the points in each local area according to the Euclidean distance between the key point and other points in the coordinate space.

Specifically, the resampled key point skeleton point cloud Each point in is used as a reference point cloud/> and distorted point cloud/> The local area center of /> No./> in key points/> , this method uses/> and/> Euclidean distance pairs of corresponding geometric parts in/> Neighbors of points are clustered, and the formula is as follows:

in, and/> Represents key points/> is the center of the ball,/> is the reference for the radius and the neighborhood group of the distorted point cloud;/> Represents the set of all points in the key point skeleton of the reference point cloud;/> Represents the set of all points in the key point skeleton of the distorted point cloud;/> represents the square of the second norm.

S103, respectively extract the three-dimensional gradient amplitude features and the three-dimensional gradient direction map features of the reference point cloud and the distorted point cloud in the local area, and calculate the three-dimensional gradient amplitude similarity of the reference point cloud and the distorted point cloud based on the three-dimensional gradient amplitude features. The three-dimensional gradient pattern feature calculates the three-dimensional gradient pattern similarity between the reference point cloud and the distorted point cloud, and calculates the joint three-dimensional gradient feature similarity based on the three-dimensional gradient amplitude similarity and the three-dimensional gradient pattern similarity.

Specifically, on the basis of extracting the three-dimensional gradient amplitude features, the calculation method of the eight directions of the three-dimensional gradient is expanded, the K nearest neighbor search algorithm is used to calculate the response intensity value of each point, and a local parameterized model of the three-dimensional gradient information is constructed, as follows.

For points in the local area of 3D point cloud , the three-dimensional gradient amplitude calculation formula is as follows:

in, ,/> ,/> It represents the gradient amplitude difference in three directions of points in each local neighborhood group.

The division angle formula of the eight directions of the three-dimensional gradient is as follows:

The calculation formula of its gradient direction diagram is as follows:

in, It is a point calculated from the angle divided by the eight directions of the three-dimensional gradient/> Directional convolution kernel;/> Represents the convolution operation.

Use a local parameterized model to extract the three-dimensional gradient features of the reference and distorted point clouds and calculate their feature similarities, including three-dimensional gradient amplitude features and gradient direction map features.

Specifically, based on the reference point cloud The three-dimensional gradient amplitude characteristics/> with gradient direction map , and distorted point cloud/> The three-dimensional gradient amplitude characteristics/> with gradient pattern/> , calculate the three-dimensional gradient amplitude similarity of the reference point cloud and the distorted point cloud/> Similarity to three-dimensional gradient direction ,Calculated as follows:

in, and/> is a constant used to ensure numerical stability.

This embodiment jointly considers three-dimensional gradient amplitude and direction similarity as a comprehensive index for point cloud quality evaluation. The calculation method can be expressed as a combination of them:

Among them, parameters is used to adjust/> and/> A positive number of relative importance between.

S104, based on the joint three-dimensional gradient feature similarity, use the response intensity value to perform weighted pooling to obtain the objective quality score of the distorted point cloud.

Specifically, the objective quality score of the point cloud is calculated using the three-dimensional response intensity value weighted pooling method. ,as follows:

in, and/> For the response intensity values of the reference point cloud and the distorted point cloud, the larger value of the two is selected as the weight in the three-dimensional edge intensity pooling strategy/> ;/> Represents the local range of the reference point cloud and distortion point cloud, that is, the three-dimensional point xyz falls within this local range;/> Represents the objective quality score of the distorted point cloud.

As shown in Figure 3, this embodiment also discloses a reference point cloud quality assessment system that integrates graph resampling and gradient features, including:

The key point extraction module 301 is used to extract key points from the input reference point cloud using a graph-based key point resampling method;

The local area point clustering module 302 is used to divide the local neighborhood group of the reference point cloud and the distortion point cloud with the key point as the center, and cluster each key point according to the Euclidean distance between the key point and other points in the coordinate space. Points within a local area;

The gradient feature similarity calculation module 303 is used to respectively extract the three-dimensional gradient amplitude features and the three-dimensional gradient pattern features of the reference point cloud and the distortion point cloud in the local area, and calculate the reference point cloud and the distortion point cloud based on the three-dimensional gradient amplitude feature. Three-dimensional gradient amplitude similarity. Calculate the three-dimensional gradient pattern similarity between the reference point cloud and the distorted point cloud based on the three-dimensional gradient pattern feature. Calculate the joint three-dimensional gradient feature similarity based on the three-dimensional gradient amplitude similarity and the three-dimensional gradient pattern similarity. Spend;

The quality assessment module 304 is used to perform weighted pooling based on the joint three-dimensional gradient feature similarity and use the response intensity value to obtain an objective quality score of the distorted point cloud.

A specific implementation of a reference point cloud quality assessment system that fuses graph resampling and gradient features. The same reference point cloud quality assessment method that fuses graph resampling and gradient features will not be repeated in this embodiment.

The above are only specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantive changes to the present invention using this concept shall constitute an infringement of the protection scope of the present invention.

Claims (7)

1. A reference point cloud quality assessment method that combines graph resampling and gradient features, which is characterized by: S101, use the graph-based key point resampling method to extract key points from the input reference point cloud; S102, divide the local neighborhood group of the reference point cloud and the distortion point cloud with the key point as the center, and cluster the points in each local area according to the Euclidean distance between the key point and other points in the coordinate space; S103, respectively extract the three-dimensional gradient amplitude features and the three-dimensional gradient direction map features of the reference point cloud and the distorted point cloud in the local area, and calculate the three-dimensional gradient amplitude similarity of the reference point cloud and the distorted point cloud based on the three-dimensional gradient amplitude features. The three-dimensional gradient pattern feature calculates the three-dimensional gradient pattern similarity between the reference point cloud and the distorted point cloud, and calculates the joint three-dimensional gradient feature similarity based on the three-dimensional gradient amplitude similarity and the three-dimensional gradient pattern similarity; S104, based on the joint three-dimensional gradient feature similarity, use the response intensity value to perform weighted pooling to obtain the objective quality score of the distorted point cloud. 2. The reference point cloud quality assessment method that fuses graph resampling and gradient features according to claim 1, characterized in that the S101 specifically includes: Reference point cloud for input , use high-pass image filtering method to // Perform resampling to extract key points; specifically use the graph transfer operator/> To describe the relative position relationship between nodes,/> Represents graph transfer operator/> dimensions; use adjacency matrix/> and transfer matrix/> to express/> , that is/> , every linear shift invariant graph filter is/> polynomial function: ; in, Represents a linear shift invariant graph filter;/> Represents the linear shift invariant graph filter./> factors, Represents the graph transfer operator /> Power;/> Indicates the first/> coefficient;/> Indicates the length of the graphics filter;/> represents the degree matrix;/> Express/> The inverse matrix;/> represents the identity matrix; Implemented using Harr-like filters The high-pass filtering is as follows: ; in, Represents the high-pass filtering of Har-like filter;/> Express/> of/> feature vector;/> Express/> eigenvalues; in the graph vertex domain, the Click/> The frequency response formula of is: ; in, Indicates the first/> Click/> The frequency response formula;/> Represents the total number of points in the spatial frequency domain;/> Indicates the first Click/> adjacent points; For input reference point cloud Extract key points to generate skeleton point cloud/> . 3. The reference point cloud quality assessment method that fuses graph resampling and gradient features according to claim 2, characterized in that said S102 specifically includes: The resampled key point skeleton point cloud Each point in is used as a reference point cloud/> and distorted point cloud/> The local area center of /> No./> in key points/> , use/> and/> Euclidean distance pairs of corresponding geometric parts in/> Neighbors of points are clustered, and the formula is as follows: ; in, Represented as key points/> is the center of the ball,/> is the neighborhood group of the reference point cloud with radius;/> Represented as key points/> is the center of the ball,/> is the neighborhood group of the distorted point cloud with radius;/> Represents the set of all points in the key point skeleton of the reference point cloud;/> Represents the set of all points in the key point skeleton of the distorted point cloud;/> represents the square of the second norm. 4. The reference point cloud quality assessment method that fuses graph resampling and gradient features according to claim 3, characterized in that said S103 specifically includes: According to the reference point cloud The three-dimensional gradient amplitude characteristics/> with three-dimensional gradient pattern/> , and distorted point cloud/> The three-dimensional gradient amplitude characteristics/> with gradient pattern/> , calculate the three-dimensional gradient amplitude similarity of the reference point cloud and the distorted point cloud/> And three-dimensional gradient direction similarity/> ,Calculated as follows: ; ; in, and/> is a constant used to ensure numerical stability; For points in the local area of the reference point cloud , three-dimensional gradient amplitude characteristics/> Calculated as follows: ; ; in, ,/> ,/> Respectively represent the gradient amplitude differences of points in the local neighborhood group of the reference point cloud in the three directions of x, y, and z;/> Represents the gradient calculation method on the xyz axis of points within the local neighborhood group of the reference point cloud; The three-dimensional gradient direction is introduced to describe the change information in the point cloud space, with key points is the center of the sphere,/> Constructs a local area for the radius, for points falling within the local area/> , the formula for calculating the spherical coordinate value of this point is as follows: ; ; in, Indicates point/> The polar angle of /> Indicates point/> Azimuth angle, perform direction clustering based on the spherical coordinate values of all points in the local area; set the eight directions of the three-dimensional gradient. The closer the spherical coordinate values of the points in the local area are to the direction, the clustering will be in that direction; the three-dimensional gradient The eight directions are calculated based on the polar angle and azimuth angle, and are equally spaced in space; the specific value calculation formula is: ; ; in ;/> Indicates polar angle/> direction,/> represents the azimuth angle direction, the two directions combine with each other to form a three-dimensional gradient eight directions; 3D gradient pattern Calculated as follows: ; in, for point/> Convolution kernel in three-dimensional eight directions. The direction of the convolution kernel depends on the direction of the three-dimensional gradient in which the point is clustered;/> Represents the convolution operation; For points in the local area of the distorted point cloud , three-dimensional gradient amplitude characteristics/> Calculated as follows: ; ; in, ,/> ,/> Respectively represent the gradient amplitude differences of points in the local neighborhood group of the distorted point cloud in the three directions of x, y, and z;/> Represents the gradient calculation method on the xyz axis of points within the local neighborhood group of the distorted point cloud; The three-dimensional gradient direction is introduced to describe the change information in the point cloud space, with key points is the center of the sphere,/> Constructs a local area for the radius, for points falling within the local area/> , the formula for calculating the spherical coordinate value of this point is as follows: ; ; in, Indicates point/> The polar angle of /> Indicates point/> Azimuth angle, perform direction clustering based on the spherical coordinate values of all points in the local area; set the eight directions of the three-dimensional gradient. The closer the spherical coordinate values of the points in the local area are to the direction, the clustering will be in that direction; the three-dimensional gradient The eight directions are calculated based on the polar angle and azimuth angle, and are equally spaced in space; the specific value calculation formula is: ; ; in ;/> Indicates polar angle/> direction,/> represents the azimuth angle direction, the two directions combine with each other to form a three-dimensional gradient eight directions; 3D gradient pattern Calculated as follows: ; in, for point/> Convolution kernel in eight three-dimensional directions. The direction of the convolution kernel depends on the direction of the three-dimensional gradient in which the point is clustered; The joint three-dimensional gradient feature similarity is calculated based on the three-dimensional gradient amplitude similarity and the three-dimensional gradient direction map similarity, as follows: ; Among them, parameters is used to adjust/> and/> A positive number of relative importance between. 5. The reference point cloud quality assessment method for fused graph resampling and gradient features according to claim 4, characterized in that S104 is as follows: ; ; in, is the response intensity value of points in the local area of the reference point cloud;/> The response intensity value of points in the local area of the distorted point cloud;/> Indicates the number of local areas;/> Represents the local range of the reference point cloud and distortion point cloud, that is, the three-dimensional point xyz falls within this local range;/> Represents the objective quality score of the distorted point cloud. 6. The reference point cloud quality assessment method that fuses graph resampling and gradient features according to claim 5, wherein the calculation method of the response intensity value is a K nearest neighbor search algorithm. 7. A reference point cloud quality assessment system that integrates graph resampling and gradient features, which is characterized by: The key point extraction module is used to extract key points from the input reference point cloud using the graph-based key point resampling method; The local area point clustering module is used to divide the local neighborhood group of the reference point cloud and the distortion point cloud with the key point as the center, and cluster each local area according to the Euclidean distance between the key point and other points in the coordinate space. Points within the region; The gradient feature similarity calculation module is used to extract the three-dimensional gradient amplitude features and the three-dimensional gradient pattern features of the reference point cloud and the distorted point cloud in the local area, and calculate the three-dimensional gradient feature of the reference point cloud and the distorted point cloud based on the three-dimensional gradient amplitude feature. Gradient amplitude similarity, calculate the three-dimensional gradient pattern similarity between the reference point cloud and the distorted point cloud based on the three-dimensional gradient pattern feature, and calculate the joint three-dimensional gradient feature similarity based on the three-dimensional gradient amplitude similarity and the three-dimensional gradient pattern similarity. ; The quality assessment module is used to obtain an objective quality score of the distorted point cloud based on the joint three-dimensional gradient feature similarity and weighted pooling using response intensity values.