Semantic feature-oriented category graph low-code-rate compression system

Abstract

The invention provides a category graph low-code-rate compression system oriented to semantic features, relates to the technical field of image processing, and is applied to the category graph low-code-rate compression system oriented to the semantic features. The category graph low-code-rate compression system oriented to the semantic features comprises the steps that a sending end carries out semantic segmentation on an input image through a semantic segmentation network to obtain a semantic segmentation graph comprising a plurality of label areas, an initial contour is extracted from the image area of the label area, each point in the initial contour carries a coordinate value and a category label value of a corresponding object, polygon fitting is carried out on the initial contour through a compression network to obtain a corresponding approximate contour, the number of points of the approximate contour is smaller than that of the points of the initial contour, and the initial contour is obtained. Encoding the approximate contour corresponding to each label area through a compression network to obtain a bit stream, and sending the bit stream to a receiving end; the receiving end decompresses the bit stream through a decompression network to obtain a decompressed approximate contour; and filling and reconstructing the decompressed approximate contour through a reconstruction network to obtain a reconstructed image.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a semantic feature-oriented category graph low-bitrate compression system. Background Art

[0002] Image is the most important information carrier in multimedia communication, and the purpose of image compression is to reduce the number of transmission bits to achieve high efficiency. Image compression has been widely studied, and many classic image compression architectures have achieved high performance in using fewer bits to store or transmit to maintain high-quality images. However, classic image compression models perform poorly in low bitrate scenarios (usually less than 0.5 bitrate). The target information cannot be well preserved using limited bits, and the semantic quality in the transmitted image is severely degraded as the bitrate decreases. Summary of the invention

[0003] In view of this, an embodiment of the present application provides a semantic feature-oriented category graph low-bitrate compression system to overcome the above problems or at least partially solve the above problems.

[0004] The first aspect of the embodiment of the present application provides a semantic feature-oriented category graph low-bitrate compression system, the system comprising a transmitting end and a receiving end, wherein the transmitting end is configured with a semantic segmentation network and a compression network in a semantic encoding and decoding network based on a semantic segmentation graph, and the receiving end is configured with a decompression network and a reconstruction network in the semantic encoding and decoding network based on the semantic segmentation graph, wherein: The transmitting end is used to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area; The transmitting end is used to encode the approximate contour corresponding to each label area through the compression network to obtain a bit stream and send it to the receiving end; The receiving end is used to decompress the received bit stream through the decompression network to obtain a decompressed approximate contour, wherein each point in the decompressed approximate contour carries a coordinate value and a category label value of a corresponding object; The receiving end is used to fill the decompressed approximate contour through the reconstruction network according to the coordinate value and the category label value of the corresponding object carried by each point in the decompressed approximate contour to obtain a plurality of reconstructed label areas; The receiving end is used to perform image reconstruction based on the reconstructed multiple label areas through the reconstruction network to obtain a reconstructed image corresponding to the input image.

[0005] Optionally, the sending end is used to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area, specifically: The transmitting end is used to perform semantic segmentation on the input image through the semantic segmentation network to obtain a semantic segmentation map, wherein the semantic segmentation map includes a plurality of label regions, each label region is used to represent an image region where an object of a category is located; The sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, wherein each point in the initial contour carries a coordinate value and a category label value of a corresponding object; The sending end is used to perform polygon fitting on the initial contour through the compression network to obtain an approximate contour corresponding to each label area, the number of points included in the approximate contour is less than the number of points included in the initial contour, and each point in the approximate contour carries a coordinate value and a category label value of the corresponding object.

[0006] Optionally, the sending end is used to perform semantic segmentation on the input image through the semantic segmentation network at the sending end, and after obtaining the semantic segmentation map, filter out the label area in the semantic segmentation map whose area is smaller than a preset threshold, so as to obtain the semantic segmentation map after filtering; The sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, including: The sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map after filtering.

[0007] Optionally, the reconstruction network includes a reconstruction model, and the reconstruction model is pre-trained using a semantic segmentation map and a decompressed semantic segmentation map.

[0008] Optionally, the sending end is used to perform polygon fitting on the initial contour through the compression network to obtain an approximate contour corresponding to each label area, specifically comprising the following steps: S1: construct a point set, and for n points contained in the initial contour of each label area of ​​the semantic segmentation map, add the first point and the nth point to the point set; S2: Generate a straight line with the first point as the starting point and the nth point as the ending point, and the points between the starting point and the ending point are a point group; S3: sequentially calculating a plurality of first distances from each point in the point group to the straight line, determining a maximum first distance, and comparing the maximum first distance with a distance threshold, wherein the maximum first distance corresponds to a critical point among the n points; S4: when the maximum first distance is greater than the distance threshold, adding the critical point to the point set; S5: Based on the critical point, respectively, a straight line is generated by taking the first point as a starting point and the critical point as an ending point, and a straight line is generated by taking the critical point as a starting point and the nth point as an ending point; S6: Execute S3 to S5 in a loop until the maximum first distance is not greater than the distance threshold, and construct an approximate contour corresponding to the label area according to multiple points in the point set.

[0009] Optionally, the sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, specifically: The transmitting end is used to obtain an initial contour including a plurality of points from the tag area by adopting a breadth-first search method or a depth-first search method; The category label value of each point included in the same initial contour is the same.

[0010] Optionally, the loss function of the semantic encoding and decoding network based on the semantic segmentation graph includes a first loss function and a second loss function, wherein; The transmitting end is further used to determine the loss value of the first loss function according to the multiple label areas in the semantic segmentation map and the reconstructed multiple label areas; The transmitting end is further used to determine the loss value of the second loss function according to the input image and the reconstructed image; The sending end is also used to update the parameters of the semantic encoding and decoding network based on the semantic segmentation map based on the loss value of the first loss function and the loss value of the second loss function.

[0011] A second aspect of an embodiment of the present application provides a semantic feature-oriented category graph low-bitrate compression method, which is applied to a semantic feature-oriented category graph low-bitrate compression system, the system comprising a transmitting end and a receiving end, wherein the transmitting end is configured with a semantic segmentation network and a compression network in a semantic encoding and decoding network based on a semantic segmentation graph, and the receiving end is configured with a decompression network and a reconstruction network in the semantic encoding and decoding network based on the semantic segmentation graph, the method comprising: The transmitting end is used to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area; The transmitting end encodes the approximate contour corresponding to each label area through the compression network to obtain a bit stream and sends it to the receiving end; The receiving end decompresses the received bit stream through the decompression network to obtain a decompressed approximate contour, wherein each point in the decompressed approximate contour carries a coordinate value and a category label value of the corresponding object; The receiving end fills the decompressed approximate contour through the reconstruction network according to the coordinate value and the category label value of the corresponding object carried by each point in the decompressed approximate contour to obtain a plurality of reconstructed label areas; The receiving end performs image reconstruction based on the reconstructed multiple label regions through the reconstruction network to obtain a reconstructed image corresponding to the input image.

[0012] According to a third aspect of an embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the low-bitrate compression method for category graphs oriented to semantic features as described in the second aspect.

[0013] According to a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the method for low-bitrate compression of category graphs oriented to semantic features as described in the second aspect is implemented.

[0014] Beneficial effects of this application: The embodiment of the present application provides a low-bitrate compression system for a category map oriented to semantic features, which is particularly applicable to a low-bitrate compression system for a category map oriented to semantic features. The low-bitrate compression system for a category map oriented to semantic features is composed of a transmitter and a receiver, wherein the transmitter is responsible for encoding, and the receiver is responsible for decoding and reconstructing the image. This method uses a semantic segmentation network and a compression network to process the input image to achieve efficient and accurate image transmission. It includes: the transmitter uses a semantic segmentation network to process the input image to generate a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map. Among them, each label area in the semantic segmentation map is used to characterize the image area where an object of a category in the semantic segmentation map is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area. Further, the transmitter encodes the approximate contour through the compression network to form a bit stream, and sends it to the receiver through the network. The receiver decodes the bit stream through the decompression network to restore the decompressed approximate contour. These approximate contours also contain the coordinate values ​​and category labels of the points. Finally, the receiver uses the reconstruction network to fill the image area according to the decompressed approximate contour to generate multiple reconstructed label areas. Based on these labeled regions, the receiver performs image reconstruction and finally obtains a reconstructed image that is similar to the original input image.

[0015] Through the technical solution of the present application, not only the amount of data that needs to be transmitted in the input image is effectively reduced, but also the accuracy and completeness of the semantic information in the reconstructed image after the image is reconstructed is guaranteed. It is particularly suitable for scenarios under low bit rates and realizes efficient and reliable image communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application.

[0017] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the description of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0018] Figure 1 It is a schematic diagram of a framework of a semantic feature-oriented low-bitrate compression system for a category graph provided by an embodiment of the present application; Figure 2 is a schematic diagram of a semantic encoding and decoding network provided in an embodiment of the present application; Figure 3It is a flowchart of a low-bitrate compression method for a category graph oriented to semantic features provided by an embodiment of the present application; Figure 4 It is a schematic diagram of the framework of a low-bitrate compression device for a category graph oriented to semantic features provided in one embodiment of the present application. DETAILED DESCRIPTION

[0019] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other.

[0020] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0021] Figure 1 is a schematic diagram of a framework of a semantic feature-oriented low-bitrate compression system for a category graph provided in an embodiment of the present application. Figure 2 It is a schematic diagram of a semantic encoding and decoding network provided in one embodiment of the present application.

[0022] like Figure 1 As shown, an embodiment of the present application provides a semantic feature-oriented category graph low-bitrate compression system, the system comprising a sending end and a receiving end, wherein the sending end is configured with a semantic segmentation network and a compression network in a semantic encoding and decoding network based on a semantic segmentation graph, and the receiving end is configured with a decompression network and a reconstruction network in the semantic encoding and decoding network based on the semantic segmentation graph, wherein: The transmitting end is used to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area; The transmitting end is used to encode the approximate contour corresponding to each label area through the compression network to obtain a bit stream and send it to the receiving end; The receiving end is used to decompress the received bit stream through the decompression network to obtain a decompressed approximate contour, wherein each point in the decompressed approximate contour carries a coordinate value and a category label value of a corresponding object; The receiving end is used to fill the decompressed approximate contour through the reconstruction network according to the coordinate value and the category label value of the corresponding object carried by each point in the decompressed approximate contour to obtain a plurality of reconstructed label areas; The receiving end is used to perform image reconstruction based on the reconstructed multiple label areas through the reconstruction network to obtain a reconstructed image corresponding to the input image.

[0023] Specifically, in the semantic feature-oriented category graph low-bitrate compression system provided in an embodiment of the present application, the system includes a transmitting end and a receiving end, the transmitting end is mainly responsible for processing and encoding, and the receiving end is mainly responsible for decoding and reconstruction, wherein the semantic feature-oriented category graph low-bitrate compression system is configured with a semantic encoding and decoding network based on a semantic segmentation graph, the semantic encoding and decoding network based on the semantic segmentation graph includes a semantic segmentation network, a compression network, a decompression network and a reconstruction network, the transmitting end is configured with a semantic segmentation network and a compression network in the semantic encoding and decoding network based on the semantic segmentation graph, and the receiving end is configured with a decompression network and a reconstruction network in the semantic encoding and decoding network based on the semantic segmentation graph.

[0024] First, the sender processes the input image through a semantic segmentation network to obtain a semantic segmentation map and the approximate contour corresponding to each label area in the semantic segmentation map. Each label area in the semantic segmentation map is used to represent the image area where an object of a category is located. The approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area.

[0025] Furthermore, after obtaining the approximate contour, the number of points used to represent the semantics of the image area is greatly reduced. Therefore, the sender can encode the approximate contour corresponding to each label area through the compression network, obtain a bit stream, and send it to the receiver. In this process, a lossless compression encoding method is adopted. Specifically, sparse representation coding can be used to compress the coordinate values ​​of the points belonging to the approximate contour and the category label values ​​of the corresponding objects. Moreover, the category label value is a scalar that can be compressed with log2[max{Ci}] bits. Because the image area in the semantic segmentation map varies in area and perimeter, different compression strategies have an impact on the metric.

[0026] Furthermore, the receiving end decompresses the bit stream sent from the sending end through a decompression network configured on the receiving end to obtain a decompressed approximate contour. Each point contained in the decompressed approximate contour carries a coordinate value and a category label value of the corresponding object. Therefore, the decompressed approximate contour is the same as the approximate contour before compression.

[0027] Furthermore, the receiving end fills the decompressed approximate contour through a reconstruction network configured on the receiving end according to the coordinate value and the category label value of the corresponding object carried by each point in the decompressed approximate contour, thereby obtaining multiple reconstructed label areas.

[0028] Finally, after obtaining the reconstructed multiple label regions, the receiving end further reconstructs the images of the reconstructed multiple label regions through a reconstruction network to obtain a reconstructed image corresponding to the input image.

[0029] Through the above embodiment, the input image can be segmented into a semantic segmentation map containing multiple label areas using an image segmentation method. The label area can represent the image area where an object of a category is located in the semantic segmentation map. The image area can be determined by its contour, and then the contour is fitted with a polygon to obtain an approximate contour containing a small number of points. The approximate contour is then encoded and sent to the receiving end for decoding to obtain a decompressed approximate contour. The label area and the image are reconstructed in sequence according to the decompressed approximate contour to obtain a reconstructed image corresponding to the input image. In this way, image transmission can be completed at a low bit rate. During the transmission process, the number of transmitted bits needs to be reduced to achieve an efficient transmission effect. The polygon fitting method can reduce the transmission of some unnecessary points, and only the points contained in the approximate contour are transmitted. While improving the transmission efficiency, the semantic fidelity between the reconstructed image reconstructed by the receiving end and the input image is also guaranteed.

[0030] Optionally, the sending end is used to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area, specifically: The transmitting end is used to perform semantic segmentation on the input image through the semantic segmentation network to obtain a semantic segmentation map, wherein the semantic segmentation map includes a plurality of label regions, each label region is used to represent an image region where an object of a category is located; The sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, wherein each point in the initial contour carries a coordinate value and a category label value of a corresponding object; The sending end is used to perform polygon fitting on the initial contour through the compression network to obtain an approximate contour corresponding to each label area, the number of points included in the approximate contour is less than the number of points included in the initial contour, and each point in the approximate contour carries a coordinate value and a category label value of the corresponding object.

[0031] Specifically, in one embodiment, the sending end processes the input image through a semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map. Each label area in the semantic segmentation map is used to represent the image area where an object of a category is located. The approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area, specifically: First, the sender uses an image splitting method based on a semantic segmentation network to extract a semantic map from the input image, thereby performing semantic segmentation on the input image to obtain a semantic segmentation map. The semantic segmentation map is divided into multiple label areas, and each label area is used to represent the image area where an object of a category is located in the semantic segmentation map. The input image is the original image that the sender needs to transmit to the receiver.

[0032] Furthermore, the image area where an object is located is represented by different contours in the semantic segmentation map. Different contours are composed of multiple points. Therefore, the sender obtains multiple pixel points on the boundary of the image area of ​​each label area contained in the semantic segmentation map, that is, Figure 2 The critical points in the image segmentation map are extracted to extract the initial contour including multiple points. The points are pixels. Different initial contours correspond to an object. Each point contained in the initial contour of each object carries the coordinate value and the category label value of the corresponding object. The category label value of each point contained in the initial contour of the same object is the same. The category label value is used to characterize the corresponding object. The coordinate value is the horizontal and vertical coordinates of the point. For example, the label value of the i-th point in the initial contour of the image region j in the semantic segmentation map is Ci, indicating that the point (pixel point) belongs to the j-th object of Ci, and the position of the i-th point can be represented by the two-dimensional horizontal and vertical coordinates (xi, yi). The initial contour of the label region j in the semantic segmentation map can be represented by a set of points {(xi(j), yi(j), Ci(j))}.

[0033] Furthermore, after obtaining the initial contour of each label area, the sender performs polygon fitting on each initial contour through a compression network, thereby reducing unnecessary points in the initial contour and retaining necessary points to obtain an approximate contour corresponding to each label area. Therefore, the points contained in the approximate contour are also pixel points, and the number of points included in the approximate contour is less than the number of points included in the initial contour. Each point in the approximate contour also carries the corresponding coordinate value and the category label value of the corresponding object. That is to say, the points contained in the approximate contour of each image area are partial points in the initial contour, and these partial points can maintain the semantics of the corresponding image area.

[0034] Optionally, the sending end is used to perform semantic segmentation on the input image through the semantic segmentation network at the sending end, and after obtaining the semantic segmentation map, filter out the label area in the semantic segmentation map whose area is smaller than a preset threshold, so as to obtain the semantic segmentation map after filtering; The sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, including: The sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map after filtering.

[0035] Specifically, in one embodiment, after obtaining the semantic segmentation map, in order to further improve the efficiency of image transmission, the label area in the semantic segmentation map whose area is smaller than a preset threshold can be filtered out to further reduce the amount of data of the transmitted bit stream, and obtain the semantic segmentation map after filtering. The filtered semantic segmentation map does not contain the label area of ​​the image area whose area is smaller than the preset threshold, wherein the size of the preset threshold can be set according to actual conditions.

[0036] Optionally, the reconstruction network includes a reconstruction model, and the reconstruction model is pre-trained using a semantic segmentation map and a decompressed semantic segmentation map.

[0037] Specifically, in one embodiment, the reconstruction network of the receiving end includes a reconstruction model, and the reconstruction model is obtained by pre-training a semantic segmentation map and a decompressed semantic segmentation map. During the training process, an image data set containing a large number of annotated semantic segmentation maps is selected, and the image data set can cover various scenes, object categories and image styles. Each semantic segmentation map in the image data set is divided into multiple label areas, and each label area corresponds to an object category in the semantic segmentation map. Each semantic segmentation map corresponds to a decompressed semantic segmentation map containing annotations. The decompressed semantic segmentation map is used to evaluate the decompressed semantic segmentation map containing multiple label areas generated by the reconstruction model. During the evaluation process, indicators such as MIoU (mean cross-linking), FWIoU (frequency weighted cross-linking) and MPA (mean pixel accuracy) can be used to evaluate the quality of the semantics contained in the decompressed semantic segmentation map containing multiple label areas generated by the reconstruction model.

[0038] Optionally, the sending end is used to perform polygon fitting on the initial contour through the compression network to obtain an approximate contour corresponding to each label area, specifically comprising the following steps: S1: construct a point set, and for n points contained in the initial contour of each label area of ​​the semantic segmentation map, add the first point and the nth point to the point set; S2: Generate a straight line with the first point as the starting point and the nth point as the ending point, and the points between the starting point and the ending point are a point group; S3: sequentially calculating a plurality of first distances from each point in the point group to the straight line, determining a maximum first distance, and comparing the maximum first distance with a distance threshold, wherein the maximum first distance corresponds to a critical point among the n points; S4: when the maximum first distance is greater than the distance threshold, adding the critical point to the point set; S5: Based on the critical point, respectively, a straight line is generated by taking the first point as a starting point and the critical point as an ending point, and a straight line is generated by taking the critical point as a starting point and the nth point as an ending point; S6: Execute S3 to S5 in a loop until the maximum first distance is not greater than the distance threshold, and construct an approximate contour corresponding to the label area according to multiple points in the point set.

[0039] Specifically, in one embodiment, the sending end performs polygon fitting on the initial contour through the compression network to obtain the approximate contour corresponding to each label area, which is specifically implemented through steps S1 to S6.

[0040] First, initialize according to the above steps. For a semantic segmentation map of size H×W, it contains N label regions, ε is the approximate threshold, and the set S is the point set used to store the points that need to be retained in the initial contour. It is necessary to perform polygon fitting on the initial contour of each label region in turn.

[0041] First, in step S1, a point set is constructed for each label region in advance. Each point set is used to retain the points that need to be used as the approximate contour of the label region. Whenever a point is determined, this point is added to the corresponding point set, and the next point is determined until all points in the initial contour corresponding to the label region are traversed. The multiple points contained in the obtained point set constitute the approximate contour of the label region. Among them, the initial contour contains n points. Before traversing each point on the boundary of the initial contour, the first point and the nth point of the initial contour are first added to the point set. In the above algorithm, the point set is S, Pstart and Pend For two variables, Pstart Indicates the starting point, Pend Indicates the end point.

[0042] Further, in step S2, in each polygon fitting process, the first point is first used as the starting point and the nth point is used as the ending point to generate a straight line, which is a straight line segment connecting the first point and the nth point, and the points between the starting point and the ending point constitute a point group, wherein, when the first point is the starting point and the nth point is the ending point, the point group is composed of the points in the n points except the first point and the nth point.

[0043] Further, in step S3, for each point in the point group, multiple first distances from multiple points in the point group to the straight line are calculated in turn (i.e., n-2 first distances are obtained), and the first distance is the maximum distance from the point in the point group to the straight line (i.e., the maximum value of the distance from the point in the point group to the first point and the distance from the point in the point group to the nth point). After obtaining the first distance of each point in the point group, the maximum first distance is determined from the multiple first distances, and the maximum first distance is compared with the distance threshold, wherein the distance threshold can be set according to actual conditions, and the point with the maximum first distance corresponds to the critical point among the n points. Further, in step S4, when the maximum first distance is greater than the distance threshold, the critical point is added to the point set, and the point with the maximum first distance greater than the distance threshold is represented as an indispensable and necessary point in the process of polygon fitting the initial contour, and therefore, needs to be added to the point set. At this time, the point set contains the first point, the nth point, and the critical point.

[0044] Further, in step S5, according to the determined critical point, a straight line is generated with the first point as the starting point and the critical point as the ending point, and a straight line is generated with the critical point as the starting point and the nth point as the ending point. The critical point can divide the n points into two point groups, the first point group (such as point group A) is a plurality of points between the critical point and the first point, and the second point group (such as point group B) is a plurality of points between the critical point and the nth point.

[0045] Further, in step S6, S3 to S5 are respectively executed in a loop until the maximum first distance is not greater than the distance threshold, and an approximate contour corresponding to the label area is constructed according to multiple points in the point set. Specifically, the first point is used as the starting point. Pstart, Critical point as termination point Pend Generate a straight line (such as line A) ， Calculate the first distance of each point in point group A to line A, and refer to steps S3 to S5 to determine the critical point in point group A, and judge whether the first distance of the critical point is greater than the distance threshold. If it is greater, the critical point in point group A is determined as the next critical point. The steps corresponding to point group B are similar to those of point group A and will not be repeated. So far, new critical points are generated in point group A and point group B respectively. When the first distance of these two critical points is greater than the distance threshold, the two new critical points are added to the point set, and the two new critical points are used to compare the starting point Pstart and end point Pend Update and execute the next cycle of step S3 to step S5 until the largest first distance is not greater than the distance threshold. Then the polygon fitting process of determining the initial contour is completed, and multiple points in the obtained point set are connected one by one to obtain the approximate contour corresponding to the label area.

[0046] Through the above embodiment, in each cycle, the farthest point from the straight line between the starting point and the end point is continuously selected, and a polygon composed of fewer vertices is gradually constructed to approximate the initial contour, and finally an approximate contour is obtained. This can greatly reduce the complexity of the initial contour and reduce the amount of data in the transmission process.

[0047] Optionally, the sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, including: Acquire an initial contour including a plurality of points from the label area by using a breadth-first search method or a depth-first search method; The category label value of each point included in the same initial contour is the same.

[0048] Specifically, in one embodiment, areas of different label images of a semantic segmentation map can be determined by different initial contours. The initial contour can be determined by obtaining an initial contour of a single pixel width and a category label value and coordinate value of each point in the initial contour by a breadth-first search method (BFS) or a depth-first search method (DFS). The category label value of each point contained in the same initial contour is the same, which represents an object of the same category.

[0049] Optionally, the loss function of the semantic encoding and decoding network based on the semantic segmentation graph includes a first loss function and a second loss function, wherein; The receiving end is used to determine the loss value of the first loss function according to the multiple label areas in the semantic segmentation map and the reconstructed multiple label areas; The receiving end is used to determine the loss value of the second loss function according to the input image and the reconstructed image; The receiving end is used to update the parameters of the semantic encoding and decoding network based on the semantic segmentation map based on the loss value of the first loss function and the loss value of the second loss function.

[0050] Specifically, in one embodiment, the semantic codec network is a deep learning model that is mainly used to process the recognition, segmentation and reconstruction of specific objects or regions in an image. It can learn complex image features to achieve accurate recognition, segmentation and reconstruction of image regions.

[0051] Therefore, the loss functions involved in the semantic encoding and decoding network include a first loss function and a second loss function.

[0052] The first loss function is used to measure the difference between the semantic segmentation map generated by the semantic codec network and the actual semantic segmentation map. Specifically, the loss value of the first loss function is calculated based on the multiple label areas in the input semantic segmentation map and the multiple label areas obtained after the semantic codec network is reconstructed. This loss function is mainly used to evaluate the segmentation effect of the semantic codec network on specific categories, ensuring that the semantic codec network can accurately identify and segment various objects in the input image.

[0053] The second loss function focuses on the similarity between the reconstructed image output by the semantic codec network and the input image. The loss value of the second loss function is calculated by comparing the difference between the input image and the reconstructed image. This loss function ensures that the semantic codec network can not only perform segmentation correctly, but also accurately reconstruct an image similar to the original input image.

[0054] Finally, by combining the loss value of the first loss function with the loss value of the second loss function, the parameters of the semantic encoding and decoding network are adjusted according to the size of the loss value. This process can use optimization algorithms such as gradient descent, and the goal is to improve the overall performance and generalization ability of the semantic encoding and decoding network by minimizing the total loss value.

[0055] Through this training method, the semantic encoder-decoder network can learn to better recognize the features of different objects in the input image and accurately reconstruct the image.

[0056] Figure 3 is a flow chart of a semantic feature-oriented low-bitrate compression method for a category graph provided in an embodiment of the present application, such as Figure 3 shown.

[0057] An embodiment of the present application provides a semantic feature-oriented category graph low-bitrate compression method, the method is applied to the semantic feature-oriented category graph low-bitrate compression system described in the above embodiment, the system includes a sending end and a receiving end, wherein the sending end is configured with a semantic segmentation network and a compression network in a semantic encoding and decoding network based on a semantic segmentation graph, and the receiving end is configured with a decompression network and a reconstruction network in the semantic encoding and decoding network based on the semantic segmentation graph, and the method includes: Step S101, the sending end processes the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area; Step S102, the transmitting end encodes the approximate contour corresponding to each label region through the compression network to obtain a bit stream and sends it to the receiving end; Step S103, the receiving end decompresses the received bit stream through the decompression network to obtain a decompressed approximate contour, wherein each point in the decompressed approximate contour carries a coordinate value and a category label value of the corresponding object; Step S103, the receiving end fills the decompressed approximate contour through the reconstruction network according to the coordinate value and the category label value of the corresponding object carried by each point in the decompressed approximate contour to obtain a plurality of reconstructed label areas; Step S104: the receiving end performs image reconstruction based on the reconstructed multiple label regions through the reconstruction network to obtain a reconstructed image corresponding to the input image.

[0058] Optionally, the step S101 includes: Step S101-1, the sending end performs semantic segmentation on the input image through the semantic segmentation network to obtain a semantic segmentation map, where the semantic segmentation map includes multiple label areas, each label area is used to represent an image area where an object of a category is located; Step S101-2, the sending end extracts an initial contour including a plurality of points from the image area of ​​each label area included in the semantic segmentation map, wherein each point in the initial contour carries a coordinate value and a category label value of a corresponding object; In step S101-3, the sending end performs polygon fitting on the initial contour through the compression network to obtain an approximate contour corresponding to each label area, wherein the number of points included in the approximate contour is less than the number of points included in the initial contour, and each point in the approximate contour carries a coordinate value and a category label value of the corresponding object.

[0059] Optionally, the method further comprises: After performing semantic segmentation on the input image through the semantic segmentation network at the sending end to obtain a semantic segmentation map, the label area in the semantic segmentation map whose area is smaller than a preset threshold is filtered out to obtain a semantic segmentation map after filtering out; The step S101 - 2 includes: extracting, by the sending end, an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map after filtering.

[0060] Optionally, the reconstruction network includes a reconstruction model, and the reconstruction model is pre-trained using a semantic segmentation map and a decompressed semantic segmentation map.

[0061] Optionally, the S101-3 specifically includes the following steps: S1: construct a point set, and for n points contained in the initial contour of each label area of ​​the semantic segmentation map, add the first point and the nth point to the point set; S2: Generate a straight line with the first point as the starting point and the nth point as the ending point, and the points between the starting point and the ending point are a point group; S3: sequentially calculating a plurality of first distances from each point in the point group to the straight line, determining a maximum first distance, and comparing the maximum first distance with a distance threshold, wherein the maximum first distance corresponds to a critical point among the n points; S4: when the maximum first distance is greater than the distance threshold, adding the critical point to the point set; S5: Based on the critical point, respectively, a straight line is generated by taking the first point as a starting point and the critical point as an ending point, and a straight line is generated by taking the critical point as a starting point and the nth point as an ending point; S6: Execute S3 to S5 in a loop until the maximum first distance is not greater than the distance threshold, and construct an approximate contour corresponding to the label area according to multiple points in the point set.

[0062] Optionally, the step S101-2 includes: Acquire an initial contour including a plurality of points from the label area by using a breadth-first search method or a depth-first search method; The category label value of each point included in the same initial contour is the same.

[0063] Optionally, the loss function of the semantic encoding and decoding network based on the semantic segmentation graph includes a first loss function and a second loss function, and the method further includes: Determining a loss value of the first loss function according to the multiple label regions in the semantic segmentation map and the reconstructed multiple label regions; Determine a loss value of the second loss function according to the input image and the reconstructed image; The parameters of the semantic encoding and decoding network based on the semantic segmentation map are updated based on the loss value of the first loss function and the loss value of the second loss function.

[0064] Figure 4 is a structural diagram of a low-bitrate compression device for a category graph oriented to semantic features provided by an embodiment of the present application, such as Figure 4 shown.

[0065] An embodiment of the present application provides a semantic feature-oriented category graph low-bitrate compression device, which is applied to a semantic feature-oriented category graph low-bitrate compression system, the semantic feature-oriented category graph low-bitrate compression system comprising a transmitting end and a receiving end, wherein the transmitting end is configured with a semantic segmentation network and a compression network in a semantic encoding and decoding network based on a semantic segmentation graph, and the receiving end is configured with a decompression network and a reconstruction network in the semantic encoding and decoding network based on the semantic segmentation graph, and the device comprises: Processing module 11, used for the sending end to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area; An encoding module 12, configured for the transmitting end to encode the approximate contour corresponding to each label region through the compression network, obtain a bit stream and send it to the receiving end; A decompression module 13, configured for the receiving end to decompress the received bit stream through the decompression network to obtain a decompressed approximate contour, wherein each point in the decompressed approximate contour carries a coordinate value and a category label value of a corresponding object; The label region reconstruction module 14 is used for the receiving end to fill the decompressed approximate contour through the reconstruction network according to the coordinate value and the category label value of the corresponding object carried by each point in the decompressed approximate contour to obtain a plurality of reconstructed label regions; The image reconstruction module 15 is used for the receiving end to perform image reconstruction based on the reconstructed multiple label regions through the reconstruction network to obtain a reconstructed image corresponding to the input image.

[0066] Optionally, the processing module 11 includes: A segmentation module, used for the sending end to perform semantic segmentation on the input image through the semantic segmentation network to obtain a semantic segmentation map, wherein the semantic segmentation map includes a plurality of label regions, each label region is used to represent an image region where an object of a category is located; An initial contour extraction module, used for the sending end to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, wherein each point in the initial contour carries a coordinate value and a category label value of a corresponding object; A polygon fitting module, used for the sending end to perform polygon fitting on the initial contour through the compression network to obtain an approximate contour corresponding to each label area, wherein the number of points included in the approximate contour is less than the number of points included in the initial contour, and each point in the approximate contour carries a coordinate value and a category label value of the corresponding object; Optionally, the device further comprises: A filtering module is used to perform semantic segmentation on the input image through the semantic segmentation network at the sending end, and after obtaining a semantic segmentation map, filter out the label area in the semantic segmentation map whose area is smaller than a preset threshold, so as to obtain a semantic segmentation map after filtering; The initial contour extraction module comprises: The first initial contour extraction unit is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map after filtering through the sending end.

[0067] Optionally, the reconstruction network includes a reconstruction model, and the reconstruction model is pre-trained using a semantic segmentation map and a decompressed semantic segmentation map.

[0068] Optionally, the device further comprises: a polygon fitting unit, wherein the polygon fitting unit is configured to perform the following steps: S1: construct a point set, and for n points contained in the initial contour of each label area of ​​the semantic segmentation map, add the first point and the nth point to the point set; S2: Generate a straight line with the first point as the starting point and the nth point as the ending point, and the points between the starting point and the ending point are a point group; S3: sequentially calculating a plurality of first distances from each point in the point group to the straight line, determining a maximum first distance, and comparing the maximum first distance with a distance threshold, wherein the maximum first distance corresponds to a critical point among the n points; S4: when the maximum first distance is greater than the distance threshold, adding the critical point to the point set; S5: Based on the critical point, respectively, a straight line is generated by taking the first point as a starting point and the critical point as an ending point, and a straight line is generated by taking the critical point as a starting point and the nth point as an ending point; S6: Execute S3 to S5 in a loop until the maximum first distance is not greater than the distance threshold, and construct an approximate contour corresponding to the label area according to multiple points in the point set.

[0069] Optionally, the initial contour extraction module includes: A second initial contour extraction unit, configured to obtain an initial contour including a plurality of points from the label area by adopting a breadth-first search method or a depth-first search method; The category label value of each point included in the same initial contour is the same.

[0070] Optionally, the loss function of the semantic encoding and decoding network based on the semantic segmentation graph includes a first loss function and a second loss function, and the device further includes: A first determination module, configured to determine a loss value of the first loss function according to the multiple label regions in the semantic segmentation map and the reconstructed multiple label regions; A second determination module, used to determine a loss value of the second loss function according to the input image and the reconstructed image; An updating module is used to update the parameters of the semantic encoding and decoding network based on the semantic segmentation map based on the loss value of the first loss function and the loss value of the second loss function.

[0071] Based on the same inventive concept, another embodiment of the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the semantic feature-oriented category graph low-bitrate compression method as described in any of the above embodiments.

[0072] Based on the same inventive concept, another embodiment of the present application further provides a computer program product, including a computer program, which is executed by a processor to implement the semantic feature-oriented category graph low-bitrate compression method as described in any of the above embodiments.

[0073] Based on the same inventive concept, another embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the low-bitrate compression method for a category graph oriented to semantic features as described in any of the above embodiments is implemented.

[0074] As for the apparatus and method, since they are basically similar to the system embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the system embodiments.

[0075] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0076] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, devices, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0077] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0078] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0079] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0080] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0081] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0082] The above is a detailed introduction to a semantic feature-oriented category graph low-bitrate compression system provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, based on the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A semantic feature-oriented category graph low-bitrate compression system, characterized in that: The system includes a transmitting end and a receiving end, wherein the transmitting end is configured with a semantic segmentation network and a compression network in a semantic encoding and decoding network based on a semantic segmentation graph, and the receiving end is configured with a decompression network and a reconstruction network in the semantic encoding and decoding network based on the semantic segmentation graph, wherein: The transmitting end is used to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area; The transmitting end is used to encode the approximate contour corresponding to each label area through the compression network to obtain a bit stream and send it to the receiving end; The receiving end is used to decompress the received bit stream through the decompression network to obtain a decompressed approximate contour, wherein each point in the decompressed approximate contour carries a coordinate value and a category label value of a corresponding object; The receiving end is used to fill the decompressed approximate contour through the reconstruction network according to the coordinate value and the category label value of the corresponding object carried by each point in the decompressed approximate contour to obtain a plurality of reconstructed label areas; The receiving end is used to perform image reconstruction based on the reconstructed multiple label areas through the reconstruction network to obtain a reconstructed image corresponding to the input image.

2. The semantic feature-oriented category graph low bit rate compression system according to claim 1, characterized in that: The transmitting end is used to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area, specifically: The transmitting end is used to perform semantic segmentation on the input image through the semantic segmentation network to obtain a semantic segmentation map, wherein the semantic segmentation map includes a plurality of label regions, each label region is used to represent an image region where an object of a category is located; The sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, wherein each point in the initial contour carries a coordinate value and a category label value of a corresponding object; The sending end is used to perform polygon fitting on the initial contour through the compression network to obtain an approximate contour corresponding to each label area, the number of points included in the approximate contour is less than the number of points included in the initial contour, and each point in the approximate contour carries a coordinate value and a category label value of the corresponding object.

3. The semantic feature-oriented category graph low bit rate compression system according to claim 2, characterized in that: in: The transmitting end is used to perform semantic segmentation on the input image through the semantic segmentation network at the transmitting end, and after obtaining the semantic segmentation map, filter out the label area in the semantic segmentation map whose area is smaller than a preset threshold, so as to obtain the semantic segmentation map after filtering out; The sending end is used to extract an initial contour including multiple points from the image area of ​​each label area contained in the semantic segmentation map, including: the sending end is used to extract an initial contour including multiple points from the image area of ​​each label area contained in the semantic segmentation map after filtering.

4. The semantic feature-oriented category graph low bit rate compression system according to claim 1, characterized in that: The reconstruction network includes a reconstruction model, and the reconstruction model is obtained by pre-training using a semantic segmentation map and a decompressed semantic segmentation map.

5. The semantic feature-oriented category graph low bit rate compression system according to claim 2, characterized in that: The transmitting end is used to perform polygon fitting on the initial contour through the compression network to obtain an approximate contour corresponding to each label area, specifically comprising the following steps: S1: construct a point set, and for n points contained in the initial contour of each label area of ​​the semantic segmentation map, add the first point and the nth point to the point set; S2: Generate a straight line with the first point as the starting point and the nth point as the ending point, and the points between the starting point and the ending point are a point group; S3: sequentially calculating a plurality of first distances from each point in the point group to the straight line, determining a maximum first distance, and comparing the maximum first distance with a distance threshold, wherein the maximum first distance corresponds to a critical point among the n points; S4: when the maximum first distance is greater than the distance threshold, adding the critical point to the point set; S5: Based on the critical point, respectively, a straight line is generated by taking the first point as a starting point and the critical point as an ending point, and a straight line is generated by taking the critical point as a starting point and the nth point as an ending point; S6: Execute S3 to S5 in a loop until the maximum first distance is not greater than the distance threshold, and construct an approximate contour corresponding to the label area according to multiple points in the point set.

6. The semantic feature-oriented category graph low bit rate compression system according to claim 2, characterized in that: The sending end is used to extract an initial contour including a plurality of points from an image area of ​​each label area included in the semantic segmentation map, specifically: The transmitting end is used to obtain an initial contour including a plurality of points from the tag area by adopting a breadth-first search method or a depth-first search method; The category label value of each point included in the same initial contour is the same.

7. The semantic feature-oriented category graph low bit rate compression system according to claim 1, characterized in that: The loss function of the semantic encoding and decoding network based on the semantic segmentation graph includes a first loss function and a second loss function, wherein; The receiving end is further used to determine the loss value of the first loss function according to the multiple label areas in the semantic segmentation map and the reconstructed multiple label areas; The receiving end is further used to determine the loss value of the second loss function according to the input image and the reconstructed image; The receiving end is also used to update the parameters of the semantic encoding and decoding network based on the semantic segmentation map based on the loss value of the first loss function and the loss value of the second loss function.

8. A semantic feature-oriented low-bitrate compression method for category graphs, characterized in that: A low-bitrate compression system for a category graph oriented to semantic features, the system comprising a transmitting end and a receiving end, wherein the transmitting end is configured with a semantic segmentation network and a compression network in a semantic encoding and decoding network based on a semantic segmentation graph, and the receiving end is configured with a decompression network and a reconstruction network in the semantic encoding and decoding network based on the semantic segmentation graph, and the method comprises: The transmitting end is used to process the input image through the semantic segmentation network to obtain a semantic segmentation map and an approximate contour corresponding to each label area in the semantic segmentation map, wherein each label area in the semantic segmentation map is used to represent the image area where an object of a category is located, and the approximate contour corresponding to each label area is obtained by polygon fitting the initial contour corresponding to the label area; The transmitting end encodes the approximate contour corresponding to each label area through the compression network to obtain a bit stream and sends it to the receiving end; The receiving end decompresses the received bit stream through the decompression network to obtain a decompressed approximate contour, wherein each point in the decompressed approximate contour carries a coordinate value and a category label value of the corresponding object; The receiving end fills the decompressed approximate contour through the reconstruction network according to the coordinate value and the category label value of the corresponding object carried by each point in the decompressed approximate contour to obtain a plurality of reconstructed label areas; The receiving end performs image reconstruction based on the reconstructed multiple label regions through the reconstruction network to obtain a reconstructed image corresponding to the input image.

9. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the semantic feature-oriented category graph low-bitrate compression method according to claim 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, wherein when the computer program is executed by a processor, the method for low-bitrate compression of a category graph oriented to semantic features as claimed in claim 8 is implemented.

Images