Lightweight Video Coding Method and System Based on Image Texture and Gradient Features

Abstract

The present invention discloses a light-load video coding method and system based on image texture and gradient features, which relates to the technical field of image management. The picture before compression processing is segmented, the image texture features of each unit area after segmentation are extracted, the gradient of the texture features between different unit areas is obtained, and the maximum gradient between the unit area and the adjacent area is used as the correlation feature between the unit area and the adjacent area. The correlation features in the picture are extracted through a convolutional network to obtain the feature vectors of each unit area in the picture. After the picture is compressed, the texture features of the compressed picture are evaluated and compared with the feature vectors obtained before compression, and the unit areas with abnormal gradients are screened out for sharpening compensation, so as to improve the file quality of the compressed picture while controlling the file size.

Description

Technical Field

[0001] The present invention relates to the technical field of image management, and specifically to a lightweight video coding method and system based on image texture and gradient features. Background Art

[0002] With the increasing applications of image recording devices such as cameras and video cameras in industrial and life scenarios, and the continuous increase in the volume of a single image, new challenges are posed to image storage. Therefore, in existing common solutions, it is necessary to compress the original image recording before storage to reduce the storage space of image information.

[0003] However, after image compression, some image detail information is discarded, and there is a problem of detail loss caused by image blurring. In application scenarios with extremely high requirements for image details such as industrial inspection, it is more likely to lead to the omission of key information, thereby affecting production decision-making and quality assessment. For images with blurred details, they usually need to be discarded or detail enhancement is performed. In the prior art, the method for detail enhancement needs to process the entire image, and all image coding information of the image needs to be processed in the computer system, resulting in an operating burden on the image processing system. Summary of the Invention

[0004] The purpose of the present invention is to provide a lightweight video coding method and system based on image texture and gradient features to solve the problems proposed in the prior art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A lightweight video coding method based on image texture and gradient features, the method comprising:

[0006] Step S100: Segment the picture before compression processing, and extract the image texture features of each unit area after segmentation;

[0007] Step S200: Obtain the gradient of the texture features between different unit areas, and use the maximum gradient between the unit area and the adjacent area as the association feature between the unit area and the adjacent area;

[0008] Step S300: Extract the association features in the picture through a convolutional network to obtain the feature vectors of each unit area in the picture;

[0009] Step S400: After the picture is compressed, perform texture feature evaluation on the compressed picture, compare it with the feature vectors obtained before compression, and screen out the unit areas with abnormal gradients for sharpening compensation.

[0010] Further, step S100 includes:

[0011] Step S101: Obtain the picture before compression processing, denote the picture as the first target picture, divide the first target picture into N unit regions, and extract the gray - scale feature values and texture feature values of each unit region;

[0012] Step S102: Compose the gray - scale feature values of each unit region into a gray - scale feature vector, and compose the texture feature values of each unit region into a texture feature vector;

[0013] Step S103: Concatenate the gray - scale feature vector and the texture feature vector of each unit region into a region feature vector corresponding to each unit region;

[0014] Step S104: Calculate the gray - scale feature values of each unit region, and the gray - scale feature values include calculating the contrast of the gray - scale image, the uniformity of the gray - scale distribution, the similarity of the gray - scale distribution, and the correlation of the gray - scale values;

[0015] Among them, the uniformity of the gray - scale distribution is calculated by the inverse difference moment of the GLCM matrix. The larger the inverse difference moment, the more consistent the gray - scale levels in the image, and the more uniform the texture. The similarity of the gray - scale distribution is calculated by the homogeneity of the GLCM matrix, and the correlation of the gray - scale values is calculated by the correlation of the GLCM matrix.

[0016] Further, step S200 includes:

[0017] Step S201: Obtain the centroid of the unit region. When the distance between the centroid of a certain unit region and the centroid of another unit region is less than the distance threshold, denote the other unit region as the adjacent unit region of the certain unit region. Gather all the adjacent unit regions of the i - th unit region among the N unit regions to form the neighborhood set Q of the i - th unit region i ;

[0018] Step S202: Obtain the j - th unit region in the neighborhood set Q, obtain the gray - scale feature value D of the i - th unit region i , the gray - scale feature value D of the j - th unit region i , the Euclidean distance dis between the centroids of the i - th unit region and the j - th unit region j , calculate the gray - scale gradient G between the i - th unit region and the j - th unit region ij , G ij , G ij = |D i - D j | / dis ij ;

[0019] Step S203: Traverse the gray-scale gradients of all unit regions in the domain set Qi and the i-th unit region, and take the unit region corresponding to the maximum gray-scale gradient as the adjacent region of the i-th unit region. The gray-scale gradient between the adjacent region and the i-th unit region is recorded as the edge weight.

[0020] Further, step S300 includes:

[0021] Step S301: Aggregate the region feature vectors of all unit regions in the first target image to obtain a node feature matrix. According to the order of the unit regions in the node feature matrix, aggregate the edge weights of all unit regions to obtain an adjacency matrix;

[0022] Step S302: Input the node feature matrix and the adjacency matrix into a graph convolutional network, and calculate the feature vectors of each unit region through the graph convolutional network. The feature vectors are recorded as the first feature vectors of the unit regions;

[0023] By describing the gray-scale features and texture features in the image slices of each unit region, and by comparing the gradient changes of the gray-scale features and texture features between the unit region and its adjacent region, the connection between the unit regions is established;

[0024] This solution deeply mines the overall gray-scale distribution and texture features of the image. On the basis of extracting the surface features of the image gray-scale and texture information, such as contrast, energy, homogeneity, and correlation, it further mines the trend of the changes in the gray-scale features and texture features between regions in the picture, providing a basis for identifying the decline in the clarity of local regions of the image in the subsequent steps.

[0025] Further, step S400 includes:

[0026] Step S401: Denote the compressed image of the first target image as the second target image. Mark the unit regions of the first target image in the compressed unit regions in the second target image, obtain the image encoding of each unit region in the second target image. Each unit region's image encoding corresponds to an image encoding set. According to the arrangement order of the unit regions in the second target image, arrange the image encoding sets to obtain the first image encoding sequence;

[0027] Step S402: Repeat steps S100 to S300 to calculate the feature vectors of each unit region in the second target image. The feature vectors are recorded as the second feature vectors of the unit regions;

[0028] Step S403: Obtain the first feature vectors and the second feature vectors of each unit region before compression, and record the distance between the first feature vector and the second feature vector as the difference degree of each unit region;

[0029] Step S404: Set the difference threshold γ, and mark the unit regions with a difference greater than the difference threshold as abnormal regions;

[0030] Step S405: Perform image sharpening on all abnormal regions in the second target image, obtain the image encoding of the abnormal regions after the image sharpening process, replace the image encoding of the corresponding unit regions in the first image encoding sequence to obtain the second image encoding sequence, and feedback the second image encoding sequence to the relevant management personnel;

[0031] Detect whether the compressed image maintains the original grayscale and texture change characteristics. When the image is compressed, the sharpness of some regions decreases, and the difference in sharpness change between different regions of the image decreases, which will cause the image to look unclear. Therefore, it is necessary to perform sharpening compensation on the regions with severe sharpness decline to keep the sharpness of the compressed image similar to that of the original image;

[0032] At the same time, the regions that need to be compensated in the image are first screened, and then the sharpness compensation is performed on the screened regions, reducing the volume of the image file after compensation. This solution further controls the volume of the compressed file on the premise of maintaining the grayscale characteristics and texture characteristics of the image before and after compression.

[0033] To better implement the above method, a lightweight video encoding system based on image texture and gradient features is also proposed. The system includes: a picture feature management module, a gradient correlation management module, a feature extraction module, and an image sharpening module;

[0034] The picture feature management module is used to segment the picture and extract the image texture features of each unit region after segmentation. The gradient correlation management module is used to obtain the gradients of the texture features between different unit regions and manage the gradient features between unit regions. The feature extraction module is used to extract the correlation features in the picture through a convolutional network and manage the feature vectors of each unit region in the picture. The image sharpening module is used to evaluate the texture features of the compressed picture, compare them with the feature vectors obtained before compression, and screen out the unit regions with abnormal gradients for sharpening compensation;

[0035] Furthermore, the picture feature management module includes: a picture segmentation unit, a grayscale feature management unit, a texture feature management unit, a regional feature vector management unit, and a grayscale feature value management unit. The picture segmentation unit is used to segment the picture and manage the unit regions in the picture. The grayscale feature management unit is used to extract and manage the grayscale features of the picture. The texture feature management unit is used to extract and manage the texture features of the picture. The regional feature vector management unit is used to manage the regional feature vectors of each unit region. The grayscale feature value management unit is used to manage the grayscale feature values of each unit region;

[0036] Further, the gradient correlation management module includes: a domain management unit, a grayscale gradient management unit, and an adjacent area management unit. The domain management unit is used to perform domain matching on each unit area. The grayscale gradient management unit is used to calculate the texture gradients of the unit area and each unit area in the domain set. The adjacent area management unit is used to manage the adjacent areas of each unit area and calculate the weights of the unit area and the corresponding adjacent areas;

[0037] Further, the feature extraction module includes: a feature matrix management unit and a graph convolution unit;

[0038] The feature matrix management unit is used to manage the node feature matrix and the adjacent feature matrix. The graph convolution unit is used to extract the feature vectors of each unit area through a graph convolution network;

[0039] Further, the image sharpening module includes: a feature vector management unit, a difference comparison unit, a difference evaluation unit, and a sharpening feedback unit;

[0040] The feature vector management unit is used to manage the first feature vector and the second feature vector of each unit area. The difference comparison unit is used to manage the difference between the first feature vector and the second feature vector. The difference evaluation unit is used to mark the unit area with a difference degree greater than the difference degree threshold as an abnormal area. The sharpening feedback unit is used to sharpen the abnormal area and feedback the second image coding sequence to relevant management personnel.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. Precise feature analysis: Describe in detail the grayscale and texture features of the image slices, and establish connections by comparing the feature gradient changes between regions. The present invention can accurately grasp the overall and local features of the image;

[0043] 2. Image quality preservation: Detect the grayscale and texture change features of the compressed image, perform sharpening compensation on the areas with serious sharpness decline, and can effectively maintain the sharpness of the compressed image similar to that of the original image, ensuring the clarity of the image in terms of visual perception and improving the user experience;

[0044] 3. File size control: First screen the areas in the image that need compensation, and then perform sharpening compensation. While maintaining the image features, the file size of the compensated image is significantly reduced, achieving a balanced optimization of image quality and file size, which is beneficial for storage and transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic structural diagram of the light-load video coding system based on image texture and gradient features of the present invention;

[0046] Figure 2Schematic diagram of the process of the light-load video encoding method based on image texture and gradient features of the present invention. Detailed implementation manners

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] Embodiment: As Figure 1 and Figure 2 shown, the present invention provides a technical solution, a light-load video encoding method and system based on image texture and gradient features;

[0049] The method includes:

[0050] Step S100: Segment the picture before compression processing, and extract the image texture features of each segmented unit area;

[0051] Among them, step S100 includes:

[0052] Step S101: Obtain the picture before compression processing, record the picture as the first target picture, segment the first target picture into N unit areas, and extract the gray feature value and texture feature value of each unit area;

[0053] Step S102: Form the gray feature values of each unit area into a gray feature vector, and form the texture feature values of each unit area into a texture feature vector;

[0054] Step S103: Concatenate the gray feature vector and texture feature vector of each unit area into a region feature vector corresponding to each unit area;

[0055] Step S104: Calculate the gray feature value of each unit area, and the gray feature value includes calculating the contrast of the gray image, the uniformity of the gray distribution, the similarity of the gray distribution, and the correlation of the gray values;

[0056] In the embodiment, first, obtain the gray image of the first target picture, quantize the gray levels in the gray image into L quantization intervals to generate the GLCM matrix of the gray image, and the row index and column index of the GLCM matrix are both 0 to L - 1; [[ID=3,8]]

[0057] Calculate the contrast, energy, homogeneity, and correlation of each unit area through the GLCM matrix;

[0058] Collect the 4 feature values in the unit area to form a 4-dimensional feature vector

[0059] Calculate the LBP features in the unit area, capture the texture features of the unit area, output an n-dimensional feature vector, and splice it with the above-mentioned 4-dimensional feature vector to obtain an n+4-dimensional regional feature vector. Preferably, when the default 8-neighborhood pixel radius is 1, the LBP feature outputs a vector with n = 10 dimensions, and the regional feature vector at this time is 14 dimensions;

[0060] In the embodiment, the method for calculating the gray feature value is to select one or more items from contrast, energy, homogeneity, and correlation for weighted average to obtain the gray feature value corresponding to the unit area.

[0061] Step S200: Obtain the gradient of the texture features between different unit areas, and use the maximum gradient between the unit area and the adjacent area as the association feature between the unit area and the adjacent area.

[0062] Among them, step S200 includes:

[0063] Step S201: Obtain the centroid of the unit area. When the distance between the centroid of a certain unit area and the centroid of another unit area is less than the distance threshold, record the other unit area as the adjacent unit area of the certain unit area, and gather all the adjacent unit areas of the i-th unit area among the N unit areas to form the neighborhood set Q of the i-th unit area i ;

[0064] Step S202: Obtain the j-th unit area in the neighborhood set Q i , obtain the gray feature value D i of the i-th unit area, the gray feature value D j of the j-th unit area, the Euclidean distance dis ij between the centroid of the i-th unit area and the centroid of the j-th unit area, and calculate the gray gradient G ij between the i-th unit area and the j-th unit area, G ij = |D i - D j | / dis ij ;

[0065] Step S203: Traverse the gray gradients between all the unit areas in the neighborhood set Qi and the i-th unit area, and use the unit area corresponding to the maximum gray gradient as the adjacent area of the i-th unit area. The gray gradient between the adjacent area and the i-th unit area is recorded as the edge weight;

[0066] In the embodiment, in order to prevent the calculation result from overflowing due to the denominator approaching 0 during the calculation of the gray gradient, G ijThe second formula, G ij =|D i -D j | / (dis ij +ε), where ε represents the remainder, and the preferred ε=e -5 , e is a natural constant.

[0067] Step S300: extracting the associated features in the image through a convolutional network to obtain the feature vectors of each unit area in the image;

[0068] Wherein, step S300 includes:

[0069] Step S301: Collecting the regional feature vectors of all unit regions in the first target image to obtain a node feature matrix, and collecting the edge weights of all unit regions according to the order of the unit regions in the node feature matrix to obtain an adjacency matrix;

[0070] Step S302: Input the node feature matrix and the adjacency matrix into the graph convolutional network, and calculate the eigenvector of each unit area through the graph convolutional network. The eigenvector is recorded as the first eigenvector of the unit area;

[0071] In this embodiment, the first target image is divided into 400 unit areas, and a 14-dimensional regional feature vector is extracted from each unit area to form a node feature matrix X with 400 rows and 14 columns. The row order of the node feature matrix is obtained, and the edge weights of the adjacent areas of each unit area are obtained in order. The weight of the non-adjacent area is 0. All edge weights and weight values are collected to obtain the adjacency matrix A.

[0072] Calculate the matrix B, B=D -1 / 2 (A+I)D -1 / 2 , where D is the degree matrix, representing the adjacency relationship of the unit area, and I is the identity matrix;

[0073] Extract feature vectors through two layers of graph convolutional layers:

[0074] G1=ReLU(BXW0), where ReLU represents the activation function, W0 is the weight matrix, which is the learnable parameter matrix. The dimension of W0 is 14×64, and the output dimension of G1 is 400×64.

[0075] G2=ReLU(BG1W1), W1 is the weight matrix, which is a learnable parameter matrix with a dimension of 64×128.

[0076] Step S400: After the image is compressed, texture features of the compressed image are evaluated and compared with feature vectors obtained before compression, and unit areas with abnormal gradients are screened out for sharpening compensation;

[0077] Among them, step S400 includes:

[0078] Step S401: Denote the compressed image of the first target image as the second target image. Mark the unit area of the first target image in the compressed unit area in the second target image, obtain the image code of each unit area in the second target image, and each unit area's image code corresponds to an image code set. Arrange the image code sets in the order of the unit areas in the second target image to obtain the first image code sequence;

[0079] Step S402: Repeat steps S100 to S300 to calculate the feature vectors of each unit area in the second target image, and denote the feature vectors as the second feature vectors of the unit areas;

[0080] Step S403: Obtain the first feature vector and the second feature vector of each unit area before compression, and denote the distance between the first feature vector and the second feature vector as the difference degree of each unit area;

[0081] Step S404: Set the difference degree threshold γ, and denote the unit areas with a difference degree greater than the difference degree threshold as abnormal areas;

[0082] Step S405: Perform image sharpening on all abnormal areas in the second target image, obtain the image codes of the abnormal areas after the image sharpening process, replace the image codes of the corresponding unit areas in the first image code sequence to obtain the second image code sequence, and feedback the second image code sequence to the relevant management personnel;

[0083] In the embodiment, number the unit areas in the first target image. After the first target image is compressed into the second target image, correspond each unit area in the first target image to the second target image, and continue to use the numbers in the first target image in the second target image;

[0084] Obtain the first feature vector M of the k-th unit area k1 and the second feature vector M k2 , by calculating the second-order norm of M k1 and M k2 , obtain the distance ΔM between the first feature vector and the second feature vector k ;

[0085] ΔM k =||M k1 -M k2 ||2 to calculate the global difference degree M global , M global =(ΔM1 + ΔM2 + ΔM3 + …… + ΔM N ) / N, where N represents the total number of unit areas;

[0086] Calculate the distances between the first eigenvectors and the second eigenvectors of all unit regions one by one, and calculate the standard deviation σ of the distances. The difference threshold γ = M global -3σ;

[0087] When outputting the abnormal regions, a visualization scheme such as a heat map or a bar chart can be adopted to represent the amount by which the difference degree of the abnormal regions exceeds the difference threshold.

[0088] The system includes: a picture feature management module, a gradient correlation management module, a feature extraction module, and an image sharpening module;

[0089] Among them, the picture feature management module is used to segment the picture and extract the image texture features of each unit region after segmentation. The picture feature management module includes: a picture segmentation unit, a grayscale feature management unit, a texture feature management unit, a region feature vector management unit, and a grayscale feature value management unit. Among them, the picture segmentation unit is used to segment the picture and manage the unit regions in the picture. The grayscale feature management unit is used to extract and manage the grayscale features of the picture. The texture feature management unit is used to extract and manage the texture features of the picture. The region feature vector management unit is used to manage the region feature vectors of each unit region. The grayscale feature value management unit is used to manage the grayscale feature values of each unit region;

[0090] Among them, the gradient correlation management module is used to obtain the gradients of the texture features between different unit regions and manage the gradient features between unit regions. Among them, the gradient correlation management module includes: a domain management unit, a grayscale gradient management unit, and an adjacent region management unit. Among them, the domain management unit is used to perform domain matching for each unit region. The grayscale gradient management unit is used to calculate the texture gradients of the unit region and each unit region in the domain set. The adjacent region management unit is used to manage the adjacent regions of each unit region and calculate the weights of the unit region and the corresponding adjacent regions;

[0091] Among them, the feature extraction module is used to extract the associated features in the picture through a convolutional network and manage the feature vectors of each unit region in the picture. Among them, the feature extraction module includes: a feature matrix management unit and a graph convolution unit. The feature matrix management unit is used to manage the node feature matrix and the adjacent feature matrix. The graph convolution unit is used to extract the feature vectors of each unit region through a graph convolutional network;

[0092] Among them, the image sharpening module is used to evaluate the texture features of the compressed picture, compare with the feature vectors obtained before compression, and screen out the unit regions with abnormal gradients for sharpening compensation. Among them, the image sharpening module includes: a feature vector management unit, a difference comparison unit, a difference evaluation unit, and a sharpening feedback unit. Among them, the feature vector management unit is used to manage the first feature vector and the second feature vector of each unit region, the difference comparison unit is used to manage the difference between the first feature vector and the second feature vector, the difference evaluation unit is used to mark the unit regions with a difference degree greater than the difference degree threshold as abnormal regions, and the sharpening feedback unit is used to sharpen the abnormal regions and feedback the second image coding sequence to relevant management personnel.

[0093] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A light-load video coding method based on image texture and gradient features, characterized in that: The method includes the steps of: Step S100: Segment the picture before compression processing, and extract the image texture features of each unit area after segmentation; Step S200: Obtain the gradients of the texture features between different unit areas, and take the maximum gradient between the unit area and the adjacent area as the association feature between the unit area and the adjacent area; Step S300: Extract the association features in the picture through a convolutional network to obtain the feature vectors of each unit area in the picture; Step S400: After the picture is compressed, evaluate the texture features of the compressed picture, compare them with the feature vectors obtained before compression, and screen out the unit areas with abnormal gradients for sharpening compensation; Step S400 includes: Step S401: Denote the picture after compressing the first target picture as the second target picture, mark the unit areas of the first target picture in the second target picture after compression, obtain the image encoding of each unit area in the second target picture, each unit area's image encoding corresponds to an image encoding set, and arrange the image encoding sets in the order of the unit areas in the second target picture to obtain the first image encoding sequence; Step S402: Repeat Step S100 to Step S300 to calculate the feature vectors of each unit area in the second target picture, and denote the feature vectors as the second feature vectors of the unit areas; Step S403: Obtain the first feature vectors and the second feature vectors of each unit area before compression, and denote the distance between the first feature vector and the second feature vector as the difference degree of each unit area; Step S404: Set the difference degree threshold γ, and denote the unit areas with a difference degree greater than the difference degree threshold as abnormal areas; Step S405: Perform image sharpening on all abnormal areas in the second target picture, obtain the image encoding of the abnormal areas after the image sharpening process, replace the image encoding of the corresponding unit areas in the first image encoding sequence to obtain the second image encoding sequence, and feedback the second image encoding sequence to relevant management personnel.

2. The light-load video coding method based on image texture and gradient features according to claim 1, characterized in that: Step S100 includes: Step S101: Obtain the picture before compression processing, denote the picture as the first target picture, segment the first target picture into N unit areas, and extract the gray feature values and texture feature values of each unit area; Step S102: Form the gray feature vectors from the gray feature values of each unit area, and form the texture feature vectors from the texture feature values of each unit area; Step S103: Concatenate the gray feature vector and the texture feature vector of each unit area into the area feature vector corresponding to each unit area; Step S104: Calculate the gray feature values of each unit area, and the gray feature values include calculating the contrast of the gray image, the uniformity of the gray distribution, the similarity of the gray distribution, and the correlation of the gray values.

3. The light-load video coding method based on image texture and gradient features according to claim 2, characterized in that: Step S200 includes: Step S201: Obtain the centroid of a unit area. When the distance between the centroid of a certain unit area and the centroid of another unit area is less than a distance threshold, record the other unit area as the adjacent unit area of the certain unit area. Aggregate all the adjacent unit areas of the i-th unit area among the N unit areas to form the neighborhood set Q of the i-th unit area i ; Step S202: Obtain the domain set Q i Obtain the j-th unit region in i , and obtain the gray-scale feature value D of the i-th unit region i The gray-scale feature value D of the j-th unit region j The Euclidean distance dis between the centroid of the i-th unit region and the j-th unit region ij Calculate the gray-scale gradient G between the i-th unit region and the j-th unit region ij G ij =|D i -D j | / dis ij ; Step S203: Traverse all unit regions in the domain set Q i and the gray-level gradient between all unit regions and the i-th unit region, and take the unit region corresponding to the maximum gray-level gradient as the adjacent region of the i-th unit region. The gray-level gradient between the adjacent region and the i-th unit region is recorded as the edge weight.

4. The light-load video coding method based on image texture and gradient features according to claim 3, characterized in that: Step S300 includes: Step S301: Pool the area feature vectors of all unit areas in the first target picture to obtain a node feature matrix, and pool the edge weights of all unit areas in the order of the unit areas in the node feature matrix to obtain an adjacency matrix; Step S302: Input the node feature matrix and the adjacency matrix into the graph convolutional network, and calculate the feature vectors of each unit area through the graph convolutional network. The feature vectors are denoted as the first feature vectors of the unit areas.

5. A light-load video coding system based on image texture and gradient features, which is used to execute the light-load video coding method based on image texture and gradient features according to any one of claims 1-4, and is characterized in that: The system includes: a picture feature management module, a gradient correlation management module, a feature extraction module, and an image sharpening module; The picture feature management module is used to segment pictures and extract the image texture features of each unit area after segmentation. The gradient correlation management module is used to obtain the gradients of texture features between different unit areas and manage the gradient features between unit areas. The feature extraction module is used to extract the associated features in the picture through a convolutional network and manage the feature vectors of each unit area in the picture. The image sharpening module is used to evaluate the texture features of the compressed picture, compare them with the feature vectors obtained before compression, and screen out the unit areas with abnormal gradients for sharpening compensation.

6. The light-load video coding system based on image texture and gradient features according to claim 5, characterized in that: The picture feature management module includes: a picture segmentation unit, a grayscale feature management unit, a texture feature management unit, a regional feature vector management unit, and a grayscale feature value management unit; The picture segmentation unit is used to segment pictures and manage the unit areas in the pictures. The grayscale feature management unit is used to extract and manage the grayscale features of the pictures. The texture feature management unit is used to extract and manage the texture features of the pictures. The regional feature vector management unit is used to manage the regional feature vectors of each unit area. The grayscale feature value management unit is used to manage the grayscale feature values of each unit area.

7. The light-load video coding system based on image texture and gradient features according to claim 5, characterized in that: The gradient correlation management module includes: a neighborhood management unit, a grayscale gradient management unit, and an adjacent area management unit; The neighborhood management unit is used to perform neighborhood matching on each unit area. The grayscale gradient management unit is used to calculate the texture gradients between the unit area and each unit area in the neighborhood set. The adjacent area management unit is used to manage the adjacent areas of each unit area and calculate the weights between the unit area and the corresponding adjacent areas.

8. The light-load video coding system based on image texture and gradient features according to claim 5, characterized in that: The feature extraction module includes: a feature matrix management unit and a graph convolutional unit; The feature matrix management unit is used to manage the node feature matrix and the adjacency feature matrix. The graph convolutional unit is used to extract the feature vectors of each unit area through the graph convolutional network.

9. The light-load video coding system based on image texture and gradient features according to claim 5, wherein: The image sharpening module includes: a feature vector management unit, a difference comparison unit, a difference evaluation unit, and a sharpening feedback unit; The feature vector management unit is used to manage the first feature vectors and the second feature vectors of each unit area. The difference comparison unit is used to manage the differences between the first feature vectors and the second feature vectors. The difference evaluation unit is used to mark the unit areas with a difference degree greater than the difference degree threshold as abnormal areas. The sharpening feedback unit is used to sharpen the abnormal areas and feedback the second image coding sequence to relevant management personnel.

Images