SAR (Synthetic Aperture Radar) image high-density compression reconstruction method based on priori knowledge feature difference
Through a high-density compression and reconstruction method of SAR images based on prior knowledge feature differences, the problems of difficult image transmission and poor quality in communication-restricted environments in multi-SAR collaborative work are solved, and efficient image compression and reconstruction are achieved.
Patent Information
- Application Number
- CN202510793627.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-13
- Publication Date
- 2025-10-17
AI Technical Summary
Existing multi-SAR collaborative working methods are difficult to meet the requirements of transmission communication volume and image quality in communication-restricted environments, and there are problems such as important information loss, noise enhancement and resolution loss.
A high-density compression and reconstruction method for SAR images based on prior knowledge feature differences is adopted. By constructing an importance network, a difference network and an entropy coding network, the key feature areas and affine transformation parameters of the image are extracted and compressed. The image generation network is used for reconstruction, and the target optimization function is optimized by combining the similarity error and classification recognition error.
Significantly reduce the amount of image transmission data, improve image compression ratio, and enhance the quality of reconstructed images, thus achieving online generation of high-quality images.
Smart Images
Figure CN120807665A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of artificial intelligence technology group cooperation, and particularly relates to a SAR image high-density compression reconstruction method based on prior knowledge feature difference. BACKGROUND
[0002] Synthetic Aperture Radar (SAR) is a technology that generates high-resolution radar images by transmitting microwave energy to the ground and collecting the reflected signals. Due to the high resolution and wide field of view of SAR images, it has been widely used in many fields such as military target detection, weather prediction, land use and environmental monitoring. By using multiple SAR radar systems to work together and integrating multi-SAR system data processing capabilities, high-resolution, multi-angle and multi-modal imaging detection of target areas can be achieved, which can effectively improve the detection accuracy and reliability.
[0003] Current multi-SAR cooperative work mainly uses traditional compression encoding methods to realize image exchange transmission between systems. In the case of limited communication environment, existing methods cannot meet the requirements of transmission traffic and image quality, and face the following problems: (1) important information loss: there may be some important targets or specific areas in SAR images, such as key targets, target edges, texture details, etc., which are of great significance to target recognition and situation analysis. If these important information cannot be correctly preserved during compression, it may lead to a decrease in the accuracy of subsequent analysis and decision-making. (2) Noise enhancement: during SAR image reconstruction, noise may be introduced or the noise in the original image may be enhanced due to the limitations of the reconstruction algorithm or interference in the signal processing process. The presence of noise may reduce the clarity and visualization quality of the image. (3) Resolution loss: SAR image reconstruction usually involves interpolation or resampling of the original data. This may result in loss of image resolution, making it impossible to completely restore some details and small-scale features, thus reducing the image's detail resolution. SUMMARY
[0004] The technical problem solved by the present application is to overcome the shortcomings of the prior art and provide a SAR image high-density compression reconstruction method based on prior knowledge feature difference, which solves the problems of image data transmission difficulty and poor received image quality in the prior art under limited communication environment.
[0005] The application is achieved by the technical scheme that the application discloses a SAR image high-density compression reconstruction method based on prior knowledge feature difference, which comprises the following steps: taking a simulation SAR image containing a target as a prior knowledge image; constructing an image compression model, using the image compression model to extract a key feature region of the target in the original image and a key feature region of the target in the prior knowledge image respectively, calculating affine transformation parameters of the key feature region of the target in the original image and the key feature region of the target in the prior knowledge image, and compressing and encoding the affine transformation parameters to obtain an intermediate variable; constructing an image reconstruction model, using the image reconstruction model to decode the intermediate variable to obtain the affine transformation parameters, generating a target reconstruction image according to the prior knowledge image and the affine transformation parameters; and obtaining a target optimization function according to a similarity error and a classification recognition error between the original image and the target reconstruction image.
[0006] The SAR image high-density compression reconstruction method based on prior knowledge feature difference further comprises the following step: integrating and training the compression model and the reconstruction model according to the target optimization function.
[0007] The SAR image high-density compression reconstruction method based on prior knowledge feature difference comprises the following steps of constructing the image compression model: constructing an importance network, which is used to extract the key feature region of the target in the original image and the prior knowledge image, and filter out background features and target weakly related feature information; constructing a difference network, which is used to calculate the affine transformation parameters of the key feature region of the target in the original image and the key feature region of the target in the prior knowledge image; and constructing an entropy encoding network, which is used to compress and encode the affine transformation parameters to obtain the intermediate variable.
[0008] The importance network in the SAR image high-density compression reconstruction method based on prior knowledge feature difference is a YOLO-V7 lightweight target detection network framework, the input is an original image and a prior knowledge image with a size of 256x256, the output is the left upper corner pixel coordinates and the right lower corner pixel coordinates of the key feature region of the target, and the key feature region of the target is scaled to a dimension of 64x64.
[0009] The difference network in the SAR image high-density compression reconstruction method based on prior knowledge feature difference adopts a convolutional neural network framework, the input is the 64x64-dimension key feature region of the original image and the prior knowledge image extracted by the importance network, and the output is linear transformation parameters and translation transformation parameters of the original image and the prior knowledge image.
[0010] In the SAR image high-density compression reconstruction method based on the prior knowledge feature difference, the image reconstruction model is constructed, including: constructing an entropy decoding network, the entropy decoding network is used for decoding the intermediate variable to obtain affine transformation parameters; constructing an image generation network, the image generation network is used for generating a target reconstruction image according to a prior knowledge image and the affine transformation parameters.
[0011] In the SAR image high-density compression reconstruction method based on the prior knowledge feature difference, the image generation network is composed of a prior feature extraction network and a reconstruction network, the input is a prior knowledge image and decoded affine transformation parameters, and the output is a target reconstruction image; wherein, for the prior feature extraction network, a layer-by-layer down-sampling manner is adopted to obtain features of different resolutions, which is used to supplement information of the prior knowledge image to the reconstruction network; the reconstruction network adopts a layer-by-layer information fusion plus up-sampling manner to process information of the prior knowledge image and the decoded affine transformation parameters to obtain the target reconstruction image.
[0012] In the SAR image high-density compression reconstruction method based on the prior knowledge feature difference, the target optimization function is obtained according to a similarity error and a classification recognition error between the original image and the target reconstruction image, including: according to pixel features of the original image and the target reconstruction image, a similarity loss function L MSE is calculated; according to classification results of inputting a pre-trained target recognition model into the original image and the target reconstruction image, a classification recognition loss function L MAE is calculated; and according to the similarity loss function L MSE and the classification recognition loss function L MAE , a target optimization function Loss is calculated.
[0013] In the SAR image high-density compression reconstruction method based on the prior knowledge feature difference, the similarity loss function L MSE is obtained through the following formula:
[0014]
[0015] Wherein, m is the total number of row pixel points of the image (the original image and the target reconstruction image), n is the total number of column pixel points of the image (the original image and the target reconstruction image), x(i,j) is the pixel point of the i-th row and the j-th column of the original image, is the pixel point of the i-th row and the j-th column of the target reconstruction image, i is the row coordinate of the pixel point, and j is the column coordinate of the pixel point.
[0016] In the SAR image high-density compression reconstruction method based on the prior knowledge feature difference, the classification recognition loss function L MAE is obtained through the following formula:
[0017]
[0018] Among them, p represents the total number of target categories, s k is the classification result of the original image after the pre-trained target recognition model, is the classification result of the target reconstructed image after the pre-trained target recognition model, and k is the target category number.
[0019] In the above-mentioned SAR image high-density compression reconstruction method based on prior knowledge feature differences, the target optimization function Loss is obtained by the following formula:
[0020] Loss = L MSE +λL MAE ;
[0021] Among them, λ represents the importance of the target reconstruction image classification error loss, L MSE is the similarity loss function, L MAE Identify loss functions for classification.
[0022] Compared with the prior art, the present invention has the following beneficial effects:
[0023] (1) The present invention uses an importance network to extract key features of relevant targets in an image for compressed transmission, thereby filtering out weakly relevant feature information for target recognition tasks in the image. This can significantly reduce the amount of image transmission data between units without affecting the reconstructed image recognition effect.
[0024] (2) The present invention constructs a target simulated SAR image as prior knowledge, obtains the affine transformation relationship between the original image and the knowledge image through a convolutional network, compresses and encodes the difference feature parameters, and significantly improves the image compression ratio;
[0025] (3) The present invention utilizes the knowledge image assisted reconstruction network to perform SAR image restoration, which can effectively improve the quality of the reconstructed image and realize the online generation of high-quality images based on a small amount of information transmitted between units. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:
[0027] Figure 1 This is a flow chart of a method for high-density compression and reconstruction of SAR images based on prior knowledge feature differences provided by an embodiment of the present invention;
[0028] Figure 2 This is a model structure diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0029] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and so that the scope of the present disclosure can be completely conveyed to those skilled in the art. It should be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.
[0030] The present embodiment provides a SAR image high-density compression reconstruction method based on prior knowledge feature difference, which comprises:
[0031] The simulation SAR image containing the target is taken as the prior knowledge image;
[0032] An image compression model is constructed, the key feature area of the original image related to the target and the key feature area of the prior knowledge image related to the target are extracted by using the image compression model, the affine transformation parameters of the key feature area of the original image related to the target and the key feature area of the prior knowledge image related to the target are calculated, and the intermediate variables are obtained by compressing and encoding the affine transformation parameters;
[0033] An image reconstruction model is constructed, the intermediate variables are decoded by using the image reconstruction model to obtain the affine transformation parameters, and the target reconstruction image is generated according to the prior knowledge image and the affine transformation parameters;
[0034] The target optimization function is obtained according to the similarity error and the classification recognition error between the original image and the target reconstruction image.
[0035] The SAR image high-density compression reconstruction method based on prior knowledge feature difference further comprises: integrating and training the compression model and the reconstruction model according to the target optimization function.
[0036] Figure 1 is a flowchart of the SAR image high-density compression reconstruction method based on prior knowledge feature difference provided by the present embodiment. As shown in Figure 1 the method comprises the following steps:
[0037] Firstly, simulation SAR images containing various targets are constructed as prior knowledge.
[0038] The construction method of the prior knowledge is:
[0039] A red-blue confrontation simulation scene is built, including a red aircraft simulation model, a red SAR radar simulation model, and three-dimensional simulation models of various types of blue targets, etc. The red flight path and SAR radar detection mode are simulated, and multi-angle and multi-scale simulation SAR images of various targets are generated and stored in the prior knowledge base. The simulation image resolution is consistent with the real detection image.
[0040] Secondly, an image compression model is constructed, the key feature regions of the original image and the same category knowledge image related to the target are extracted respectively, the affine transformation parameters of the key features of the original image and the knowledge image are calculated, and the intermediate variables are obtained by compression encoding, as shown in the following formula: Figure 2
[0041] The construction process of the compression model is as follows:
[0042] (2.1) An importance network is constructed to extract the key features of the target in the SAR image and the knowledge image, and to screen out the background features and weakly related features of the target in the image;
[0043] The construction method of the importance network framework is as follows:
[0044] The YOLO-V7 lightweight target detection network framework is adopted, the SAR image with a size of 256x256 is input, the left upper corner pixel coordinates (x1, y1) and the right lower corner pixel coordinates (x2, y2) of the key feature region of the target are output, and the feature region is scaled to 64x64 dimensions.
[0045] (2.2) A difference network is constructed to calculate the affine transformation relationship between the key features of the original image and the knowledge image;
[0046] The construction method of the difference network framework is as follows:
[0047] Considering that the knowledge image is a same category target detection image from a simulated flight trajectory, and the target size and texture of the knowledge image are similar to those of the real SAR image, the SAR image can be regarded as a scaling, translation, rotation, etc. of the knowledge image. The convolutional neural network framework is adopted to learn the affine transformation relationship between the key features of the original image and the knowledge image. The input is the 64x64 key features of the SAR image and the same category knowledge image extracted by the importance network, and the output is the linear transformation parameter γ and the translation transformation parameter β of the two images.
[0048] (2.3) An entropy encoding network is constructed to compress and encode the affine transformation coefficients of the SAR image and the knowledge image into intermediate variables in the form of byte stream for transmission between units.
[0049] The construction method of the entropy encoding network framework is as follows:
[0050] The entropy coding network is used to estimate the entropy probability model of the compressed features, so that the compressed features can be encoded and decoded, a Gaussian distribution is used as a priori, and then the model is used to estimate the mean and variance of the Gaussian distribution, the probability table obtained based on the Gaussian distribution is used for encoding, and the more accurate the probability model estimation is, the smaller the code rate consumed is.
[0051] Thirdly, an image reconstruction model is constructed, and a knowledge image is introduced into the intermediate variable to assist in enhancement, and a target reconstruction image is generated online.
[0052] The construction process of the reconstruction model is as follows:
[0053] (3.1) Construct an entropy decoding network to decode the obtained intermediate variable to obtain affine transformation parameters;
[0054] The construction method of the entropy decoding network framework is as follows:
[0055] The entropy decoding network framework is symmetrical to the entropy coding network, and the obtained intermediate variable is decoded according to the entropy model parameters obtained by training.
[0056] (3.2) Construct an image generation network to generate a target reconstruction image online by using a knowledge image and affine transformation parameters.
[0057] The construction method of the image generation network framework is as follows:
[0058] The image generation network is composed of a prior feature extraction network and a reconstruction network, and the input is a knowledge image and decoded affine transformation parameters, and the output is a target reconstruction image. For the prior feature extraction network, a layer-by-layer downsampling method is used to obtain features of different resolutions, which are used to supplement the information of the knowledge image to the reconstruction network; for the reconstruction network, a layer-by-layer information fusion + upsampling method is used to utilize the features transmitted by the prior feature extraction network, increase the resolution, reduce the channel number, and fix the output channel number through 1x1 convolution in the last layer. The reconstruction network is symmetrical to the difference network, the difference network is 4 times down-sampling, and the reconstruction network is 4 times up-sampling, and finally the target reconstruction image is obtained.
[0059] Fourthly, a target optimization function is constructed by using the similarity error and classification recognition error between the original image and the reconstruction image, and the compression model and the reconstruction model are integrated and trained.
[0060] The training process of the compression model and the reconstruction model is as follows:
[0061] (4.1) According to the pixel features of the original image and the reconstruction image, the similarity loss function L MSE is calculated.
[0062] The calculation formula of the similarity loss function L MSE is as follows:
[0063]
[0064] Wherein, m is the total number of row pixel points of the image (original image and target reconstruction image), n is the total number of column pixel points of the image (original image and target reconstruction image), x(i,j) is the pixel point of the i-th row and j-th column of the original image, Y(i,j) is the pixel point of the i-th row and j-th column of the target reconstruction image, i is the row coordinate of the pixel point, and j is the column coordinate of the pixel point.
[0065] (4.2) According to the classification result of the original image and the reconstruction image inputting the pre-trained target recognition model, the classification recognition loss function L is calculated MAE ;
[0066] The calculation formula of the classification recognition loss function L MAE is as follows:
[0067]
[0068] Wherein, p represents the total number of target categories, s k is the classification result of the original image through the pre-trained target recognition model, is the classification result of the target reconstruction image through the pre-trained target recognition model, and k is the target category serial number.
[0069] (4.3) According to the similarity loss function L MAE and the classification recognition loss function L MAE , the target optimization function Loss is calculated.
[0070] The calculation formula of the target optimization function Loss is as follows:
[0071] Loss=L MSE +λL MAE
[0072] Wherein, λ represents the importance of the reconstruction image classification error loss, λ=0 represents that only the similarity of the original image and the reconstruction image is considered, and λ=1 represents that only the target recognition consistency of the original image and the reconstruction image is considered.
[0073] (4.4) Based on the image data set, the compression model and the reconstruction model are integrated and trained in a supervised learning manner.
[0074] In order to meet the requirements of image transmission and target recognition of cluster cooperative detection under the condition of limited communication, the prior knowledge image is introduced, the image compression and reconstruction model are modeled by using deep learning, and then integrated and trained.
[0075] The training and generalization ability improvement of the intelligent model need data driving, and the image sample training set is constructed by combining the simulation SAR image and the measured SAR image, and the evolution of the image compression and reconstruction model is completed.
[0076] 1) Before the model training starts, the simulation SAR images p of various targets are constructed i As prior knowledge, the SAR images x of various targets under different flight trajectories are measured and collected i The image data training set {((x i , p i ), i), i = 1, 2, 3, …} is made with the target type as a label;
[0077] 2) A fixed number of image samples are randomly extracted from the training data set for model iterative training;
[0078] 3) The measured SAR image x i and the prior knowledge graph p i are input into the compression model to obtain the intermediate variable z;
[0079] 4) The intermediate variable z and the prior knowledge graph p i are input into the reconstruction model to obtain the reconstructed image
[0080] 5) The similarity loss function L i is calculated by using the measured SAR image x MSE and the reconstructed image ;
[0081] 6) The measured SAR image x i and the reconstructed image are input into the pre-trained target recognition model respectively to obtain one-dimensional features s i and The classification recognition loss function L MAE is calculated.
[0082] 7) According to L MSE and L MAE , the target optimization function Loss is calculated for back propagation to update the network weight parameters;
[0083] 8) Steps 2) to 7) are repeated to realize the evolution training of the image compression and reconstruction model.
[0084] Based on the inter-group image data sharing ability under the condition of reducing the communication bandwidth by thousands of times, the method of the application is verified, and the test process is as follows:
[0085] 1) Set a suitable communication bandwidth range;
[0086] 2) Through the construction of image training data set, the training of compression model and reconstruction model is realized, if the model training does not converge, the parameters are adjusted and retrained until the model converges into the next step;
[0087] 3) Under the random field scene assumption, the SAR image data of different targets are measured and collected, and the verification test of the method of the embodiment is carried out;
[0088] 4) Under the same typical scene as step 3), the prior knowledge image is cancelled, and the compression model and the reconstruction model are directly used for image compression and reconstruction, and after the model training converges, the model is verified;
[0089] 5) The test results of step 3) and step 4) are statistically compared and analyzed, and it is found that the embodiment can well solve the problems that the image data cannot be transmitted at one time and the image recognition accuracy is sharply reduced after decoding in the traditional coding and decoding method under the condition that the communication bandwidth is greatly degraded.
[0090] The embodiment introduces prior knowledge image, uses deep learning model to compress and transmit the SAR measured image and online generates; uses importance network to extract the key features of the target in the image for compression transmission, and screens out the weakly related feature information of the target recognition task in the image, which can greatly reduce the image transmission data amount between units without affecting the recognition effect of the reconstructed image; the target simulation SAR image is constructed as prior knowledge, the affine transformation relationship between the original image and the knowledge image is obtained through the convolution network, the difference feature parameters are compressed and encoded, and the image compression ratio is significantly improved; the knowledge image is used to assist the reconstruction network to restore the SAR image, which can effectively improve the quality of the reconstructed image and realize the online generation of high-quality image based on a small amount of inter-unit transmission information; the measured SAR image data set is used to verify the effectiveness of the embodiment. The embodiment solves the problems of image data transmission difficulty and poor received image quality in the prior art under the limited communication environment.
[0091] Although the present application has been disclosed with reference to the preferred embodiments above, it is not intended to limit the present application, and any person skilled in the art can make possible changes and modifications to the technical solutions of the present application by using the disclosed methods and technical contents without departing from the spirit and scope of the present application. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, which does not deviate from the technical solutions of the present application, belongs to the protection scope of the technical solutions of the present application.
Claims
1. A high-density compression reconstruction method for SAR images based on prior knowledge feature differences, characterized by include: The simulated SAR image containing the target is used as a priori knowledge image; Constructing an image compression model, using the image compression model to extract key feature areas of the original image related to the target and key feature areas of the prior knowledge image related to the target, calculating affine transformation parameters of the key feature areas of the original image related to the target and the key feature areas of the prior knowledge image related to the target, and compressing and encoding the affine transformation parameters to obtain intermediate variables; Construct an image reconstruction model, use the image reconstruction model to decode the intermediate variables to obtain affine transformation parameters, and generate the target reconstructed image based on the prior knowledge image and affine transformation parameters; The target optimization function is obtained according to the similarity error and classification recognition error between the original image and the target reconstructed image.
2. The SAR image high-density compression and reconstruction method based on prior knowledge feature difference according to claim 1 is characterized in that Also includes: The compression model and the reconstruction model are integrated and trained according to the target optimization function.
3. The SAR image high-density compression and reconstruction method based on prior knowledge feature differences according to claim 1 is characterized by: Building an image compression model involves: Constructing an importance network, wherein the importance network is used to extract key feature areas related to the target in the original image and the prior knowledge image, and filter out background features and weakly related feature information of the target; Constructing a difference network for calculating affine transformation parameters of key feature regions of the original image related to the target and key feature regions of the prior knowledge image related to the target; An entropy coding network is constructed, wherein the entropy coding network is used to compress and encode the affine transformation parameters to obtain intermediate variables.
4. The method for high-density compression and reconstruction of SAR images based on prior knowledge feature differences according to claim 3, characterized in that: The importance network adopts the YOLO-V7 lightweight target detection network framework, inputs a 256×256 original image and a priori knowledge image, outputs the upper left corner pixel coordinates and the lower right corner pixel coordinates of the key feature area of the target, and scales the key feature area of the target to 64×64 dimensions.
5. The method for high-density compression and reconstruction of SAR images based on prior knowledge feature differences according to claim 4, characterized in that: The difference network adopts a convolutional neural network framework, with the input being the key feature areas of the original image and the prior knowledge image extracted by the importance network in 64×64 dimensions, and the output being the linear transformation parameters and translation transformation parameters of the original image and the prior knowledge image.
6. The SAR image high-density compression and reconstruction method based on prior knowledge feature differences according to claim 1 is characterized by: Building an image reconstruction model includes: Constructing an entropy decoding network, wherein the entropy decoding network is used to decode the intermediate variables to obtain affine transformation parameters; An image generation network is constructed, wherein the image generation network is used to generate a target reconstructed image according to a priori knowledge image and affine transformation parameters.
7. The method for high-density compression and reconstruction of SAR images based on prior knowledge feature differences according to claim 6, characterized in that: The image generation network consists of a prior feature extraction network and a reconstruction network, with the input being a prior knowledge image and decoded affine transformation parameters, and the output being a target reconstructed image. The prior feature extraction network uses a layer-by-layer downsampling method to obtain features of different resolutions, which are used to supplement the information of the prior knowledge image to the reconstruction network. The reconstruction network uses a layer-by-layer information fusion plus sampling method to process the information of the prior knowledge image and the decoded affine transformation parameters to obtain the target reconstructed image.
8. The SAR image high-density compression and reconstruction method based on prior knowledge feature differences according to claim 1 is characterized by: The target optimization function is obtained based on the similarity error and classification recognition error between the original image and the target reconstructed image: According to the pixel features of the original image and the target reconstructed image, the similarity loss function L is calculated. MSE ; According to the classification results of the pre-trained target recognition model input by the original image and the target reconstructed image, the classification recognition loss function L is calculated. MAE ; According to the similarity loss function L MSE And classification recognition loss function L MAE , calculate the target optimization function Loss.
9. The SAR image high-density compression and reconstruction method based on prior knowledge feature difference according to claim 8 is characterized in that: similarity Loss function L MSE It is obtained by the following formula: Among them, m is the total number of row pixels, n is the total number of column pixels, x(i,j) is the pixel in the i-th row and j-th column of the original image, Reconstruct the pixel at row i and column j of the target image, where i is the row coordinate of the pixel and j is the column coordinate of the pixel.
10. The SAR image high-density compression and reconstruction method based on prior knowledge feature differences according to claim 8, characterized in that: Classification recognition loss function L MAE It is obtained by the following formula: Among them, p represents the total number of target categories, s k is the classification result of the original image after the pre-trained target recognition model, is the classification result of the target reconstructed image after the pre-trained target recognition model, and k is the target category number; The objective optimization function Loss is obtained by the following formula: Loss=L MSE +λL MAE ; Among them, λ represents the importance of the target reconstruction image classification error loss, L MSE is the similarity loss function, L MAE Identify loss functions for classification.