Multi-view SAR (Synthetic Aperture Radar) target identification method, system and device and storage medium

Abstract

The invention discloses a multi-view SAR target recognition method, system and device and a storage medium, and belongs to the technical field of SAR image recognition. The method comprises the following steps: acquiring a plurality of SAR images of different visual angles, and performing spatial dimension splicing on the plurality of SAR images of different visual angles to obtain a spliced image; dividing each SAR image into a plurality of image blocks, and extracting initial features of each image block; determining a global pole in the spliced image, establishing a global polar coordinate system according to the global pole, and calculating a global polar coordinate position code of each image block; calculating a relative position score matrix according to the global polar coordinate position code of each image block; and inputting the relative position score matrix and the initial features of the image blocks into a pre-acquired Transform model, constructing a feature attention matrix of the image blocks, superposing the relative position score matrix and the feature attention matrix, and outputting a final recognition result. According to the method, the polar coordinate position codes are constructed, so that the multi-view feature space consistency is ensured, and the recognition accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of SAR image recognition, and particularly relates to a multi-view SAR target recognition method, system, device and storage medium. Background Art

[0002] Synthetic Aperture Radar (SAR) is an important active remote sensing technology, which is widely used in many fields such as earth observation, environmental monitoring, military reconnaissance and disaster assessment. Compared with optical remote sensing, the imaging principle of SAR is based on the emission and echo reception of microwaves, and can obtain high-resolution surface information under any weather conditions and day or night. Therefore, the application scenarios of SAR images are increasing, and the accuracy requirement for SAR image recognition is increasing day by day.

[0003] In the existing methods for SAR image recognition, the commonly used position encodings include absolute position and relative position. These position encoding methods cannot consider the multi-view image spatial relationship, resulting in the lack of multi-view relative position information in the image. In the existing methods for multi-view SAR image recognition, generally, single-view independent features are first extracted and then combined to obtain multi-view features. This approach severs the integrity of multi-view data, resulting in the loss of the global perception advantage in the early fusion stage and low recognition accuracy. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a multi-view SAR target recognition method, system, device and storage medium, which can accurately describe the global relative position and multi-view spatial relationship of the image and improve the recognition accuracy.

[0005] The present invention provides the following technical solutions:

[0006] In the first aspect, a multi-view SAR target recognition method is provided, including: obtaining a plurality of SAR images with different views, and splicing the plurality of SAR images with different views in the spatial dimension to obtain a spliced image;

[0007] Dividing each SAR image into a plurality of image blocks, and extracting the initial features of each image block;

[0008] Determining the global poles in the spliced image and establishing a global polar coordinate system based on the global poles, and calculating the global polar coordinate position encoding of each image block based on the global polar coordinate system;

[0009] Calculating a relative position score matrix for quantifying the relative position between image blocks according to the global polar coordinate position encoding of each image block;

[0010] Input the relative position score matrix and the initial features of each image patch into a pre-acquired Transformer model to construct a feature attention matrix of the image patches, superimpose the relative position score matrix and the feature attention matrix, and output the final recognition result.

[0011] As an alternative technical solution of the present invention, each perspective corresponds to one of the SAR images.

[0012] As an alternative technical solution of the present invention, divide each of the SAR images into a plurality of image patches, and extract the initial features of each image patch, including:

[0013] The sizes of the plurality of image patches are the same and non-overlapping;

[0014] Extract the initial features of each image patch through a convolutional neural network model.

[0015] As an alternative technical solution of the present invention, determine the global poles in the stitched image and establish a global polar coordinate system based on the global poles, and calculate the global polar coordinate positions of each image patch based on the global polar coordinate system, including:

[0016] Take the image center of any one of the SAR images as the current local pole, and establish a local polar coordinate system with the horizontal right direction as the polar axis. The local polar coordinates of each image patch are expressed as , where represents the distance from the current local pole of the i-th SAR image to the center point of the p-th image patch, represents the distance from the current local pole of the i-th SAR image to the center point of the p-th image patch and the angle with the polar axis;

[0017] In the stitched image, take the image center of any one of the SAR images as the global pole, and establish a global polar coordinate system with the direction pointing to the radar corresponding to the SAR image as the polar axis. In the global polar coordinate system, represent the distance offset of the image center of the i-th SAR image relative to the global pole as , and represent the perspective offset of the radar corresponding to the i-th SAR image relative to the global pole as ;

[0018] Combining the local polar coordinate system and the global polar coordinate system, the global polar coordinate position encoding of the p-th image patch of the i-th SAR image is expressed as:

[0019] ;

[0020] where represents the global polar coordinate position encoding of the p-th image patch in the i-th SAR image.

[0021] As an alternative technical solution of the present invention, a relative position score matrix for quantifying the relative positions between image blocks is calculated according to the global polar coordinate position encodings of the respective image blocks, including:

[0022] The global polar coordinate position encodings of any two image blocks are respectively expressed as:

[0023] ;

[0024] ;

[0025] wherein, represents the global polar coordinate position encoding of the q-th image block in the j-th SAR image, represents the distance from the current local pole of the j-th SAR image to the center point of the q-th image block, represents the distance from the current local pole of the j-th SAR image to the center point of the q-th image block and the included angle with the polar axis, represents the distance offset of the image center of the j-th SAR image relative to the global pole, represents the viewing angle offset of the corresponding radar of the j-th SAR image relative to the global pole;

[0026] Calculate the distance between two image blocks , which is expressed as:

[0027] ;

[0028] When , the distance between the two image blocks is expressed as:

[0029] ;

[0030] The distance matrix of all image blocks is expressed as:

[0031] ;

[0032] wherein, D represents the distance matrix of all image blocks, M represents the total number of image blocks, represents the distance between the p-th image block and the q-th image block, ;

[0033] Perform linear compression on each element in the distance matrix D, which is expressed as:

[0034] ;

[0035] wherein, represents the distance between two image blocks after linear compression, denotes the maximum value of the elements in the distance matrix D;

[0036] The relative position score matrix is expressed as:

[0037] ;

[0038] where denotes the relative position score matrix, denotes the compressed distance matrix, , denotes the compressed distance between the p-th image patch and the q-th image patch, denotes the weight parameter, denotes the bias parameter.

[0039] As an alternative technical solution of the present invention, input the relative position score matrix and the initial features of each image patch into a pre-obtained Transformer model to construct a feature attention matrix of the image patches, stack the relative position score matrix and the feature attention matrix, and output the final recognition result, including:

[0040] The Transformer model includes multiple feature extraction layers, and each feature extraction layer includes a multi-head attention module;

[0041] The initial features of each image patch are input into the multi-head attention module of the first layer to obtain a feature attention matrix, which is expressed as:

[0042] ;

[0043] where denotes the feature attention matrix, Q denotes the query matrix, K denotes the key matrix, T denotes the matrix transpose, denotes the vector dimension;

[0044] In the multi-head attention module of the first layer, stack the relative position score matrix and the feature attention matrix, and calculate the first layer feature matrix, which is expressed as:

[0045] ;

[0046] where denotes the first layer feature matrix, denotes the relative position score matrix, V denotes the value matrix, denotes the softmax operation;

[0047] Input the feature matrix of the current layer into the multi-head attention module of the next layer, stack the relative position score matrix, and calculate the feature matrix of the next layer until the final recognition result is output.

[0048] In a second aspect, a multi-view SAR target recognition system is provided, including: an image acquisition module, configured to acquire SAR images from several different views, and splice the SAR images from several different views in the spatial dimension to obtain a spliced image;

[0049] An image division module, configured to divide each SAR image into several image blocks, and extract the initial features of each image block;

[0050] A polar coordinate encoding module, configured to determine the global pole point in the spliced image and establish a global polar coordinate system based on the global pole point, and calculate the global polar coordinate position encoding of each image block based on the global polar coordinate system;

[0051] A matrix calculation module, configured to calculate a relative position score matrix for quantifying the relative positions between image blocks according to the global polar coordinate position encodings of each image block;

[0052] An output module, configured to input the relative position score matrix and the initial features of each image block into a pre-acquired Transformer model, construct a feature attention matrix of the image blocks, superimpose the relative position score matrix and the feature attention matrix, and output a final recognition result.

[0053] In a third aspect, a multi-view SAR target recognition device is provided, including a processor and a storage medium; the storage medium is used for storing instructions; the processor is configured to operate according to the instructions to execute the steps of the multi-view SAR target recognition method described in the first aspect.

[0054] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the steps of the multi-view SAR target recognition method described in the first aspect are implemented.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] The multi-view SAR target recognition method provided by the present invention can accurately describe the global relative position and multi-view spatial relationship of SAR images through polar coordinate position encoding, ensure the consistency of the feature space, and enhance the correlation between image blocks from different views; meanwhile, the Transformer model utilizes the initial features of multi-view SAR images to achieve effective early feature fusion, improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flowchart of the multi-view SAR target recognition method in an embodiment of the present invention;

[0058] Figure 2 is a schematic diagram of a local polar coordinate system in an embodiment of the present invention;

[0059] Figure 3 It is a schematic diagram of the global polar coordinate system in the embodiment of the present invention;

[0060] Figure 4 It is a schematic diagram of the structure of the Transformer model in the embodiment of the present invention;

[0061] Figure 5 It is the SAR image of different military targets from different perspectives in the embodiment of the present invention;

[0062] Figure 6 It is a schematic diagram of the result of the multi - perspective SAR target recognition method in the embodiment of the present invention. Specific implementation manner

[0063] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0064] Embodiment 1

[0065] This embodiment provides a multi - perspective SAR target recognition method. As Figure 1 shown, it specifically includes the following steps:

[0066] Step 1: Obtain several SAR images from different perspectives, and splice the several SAR images from different perspectives in the spatial dimension to obtain a spliced image.

[0067] In this embodiment, several SAR images from different perspectives are pre - processed (for example, amplitude correction), but image registration has not been performed yet. Each perspective corresponds to one of the SAR images. In order to effectively fuse multi - perspective features, the method of spatial dimension splicing is adopted, aiming to retain the original spatial structure characteristics of each perspective to the greatest extent.

[0068] Step 2: Divide each SAR image into several image blocks, and extract the initial features of each image block.

[0069] Specifically, each SAR image in the spliced image is divided, and the obtained several image blocks have the same size and do not overlap; the initial features of each image block are extracted through a convolutional neural network model, and the image data is converted into local features.

[0070] Step 3: Determine the global pole point in the spliced image and establish a global polar coordinate system based on the global pole point, and calculate the global polar coordinate position encoding of each image block based on the global polar coordinate system.

[0071] This embodiment adopts polar coordinate position encoding to describe the global relative position of the image block and the rotation offset of the multi - perspective image, ensuring the consistency of the multi - perspective feature space. Specifically, it includes:

[0072] 3.1. As shown in Figure 2 , take the image center of any SAR image as the current local pole , and take the horizontal right direction as the polar axis to establish a local polar coordinate system. For the p-th image block of the i-th SAR image, its local polar coordinates are expressed as , which can accurately describe the absolute position of the image block within a single-view image.

[0073] Among them, represents the distance from the current local pole of the i-th SAR image to the center point of the p-th image block, and the value range is (0, 1). represents the distance from the current local pole of the i-th SAR image to the center point of the p-th image block and the included angle with the polar axis, and the value range is [0, 2π). Take the target center of any SAR image as the pole , and set the target center to coincide with the image center.

[0074] 3.2. As shown in Figure 3 , to unify the relative position relationship of multi-view images, it is necessary to define a global polar coordinate system. In the stitched image, take the image center of any SAR image as the global pole , and take the ray pointing to the direction of the radar corresponding to the SAR image as the polar axis to establish a global polar coordinate system. To describe the relative spatial relationship of each view, define the target center offset and the view offset .

[0075] Among them, in the global polar coordinate system, the distance offset of the image center of the i-th SAR image relative to the global pole is expressed as , and the view offset of the radar corresponding to the i-th SAR image relative to the global pole is expressed as .

[0076] 3.3. Combining the local polar coordinate system and the global polar coordinate system, the global polar coordinate position encoding of the p-th image block of the i-th SAR image is expressed as:

[0077] ;

[0078] Among them, represents the global polar coordinate position encoding of the p-th image block in the i-th SAR image.

[0079] This encoding rule can not only describe the spatial distribution of image patches in a single-view image, but also accurately characterize the rotation and translation relationships between multi-view images through relative position offsets. The concept of adjacent image patches is extended from single-view to multi-view images, enhancing the correlation between image patches from different views.

[0080] Step 4: Calculate the relative position score matrix for quantifying the relative positions between image patches based on the global polar coordinate position encodings of each image patch.

[0081] The global polar coordinate position encodings of any two image patches are respectively represented as:

[0082] ;

[0083] ;

[0084] where represents the global polar coordinate position encoding of the q-th image patch in the j-th SAR image, represents the distance from the current local pole of the j-th SAR image to the center point of the q-th image patch, represents the distance from the current local pole of the j-th SAR image to the center point of the q-th image patch and the angle with the polar axis, represents the distance offset of the image center of the j-th SAR image relative to the global pole, represents the viewing angle offset of the corresponding radar of the j-th SAR image relative to the global pole.

[0085] Calculate the distance between two image patches , which is represented as:

[0086] ;

[0087] When , the distance between the two image patches is represented as:

[0088] ;

[0089] The distance matrix of all image patches is represented as:

[0090] ;

[0091] where D represents the distance matrix of all image patches, M represents the total number of image patches, represents the distance between the p-th image patch and the q-th image patch, .

[0092] Perform linear compression on each element in the distance matrix D, linearly compress it to [-1, 1], which is represented as:

[0093] ;

[0094] Wherein, represents the distance between two linearly compressed image patches, represents the maximum value of the elements in the distance matrix D;

[0095] The relative position score matrix is expressed as:

[0096] ;

[0097] Wherein, represents the relative position score matrix, represents the compressed distance matrix, , represents the compressed distance between the p-th image patch and the q-th image patch, represents the weight parameter, represents the bias parameter.

[0098] Step Five: Input the relative position score matrix and the initial features of each image patch into a pre-obtained Transformer model, construct a feature attention matrix of the image patches, superimpose the relative position score matrix and the feature attention matrix, and output the final recognition result.

[0099] As Figure 4 shown, the Transformer model includes multiple feature extraction layers, and each feature extraction layer includes a multi-head attention module.

[0100] The initial features of each image patch are input into the multi-head attention module of the first layer to obtain a feature attention matrix, which is expressed as:

[0101] ;

[0102] Wherein, represents the feature attention matrix, Q represents the query matrix, K represents the key matrix, T represents the matrix transpose, represents the vector dimension.

[0103] In the multi-head attention module of the first layer, the relative position score matrix and the feature attention matrix are superimposed, and the first-layer feature matrix is calculated, which is expressed as:

[0104] ;

[0105] Wherein, represents the first-layer feature matrix, represents the relative position score matrix, V represents the value matrix, represents the softmax operation.

[0106] After the obtained first-layer feature matrix undergoes non-linear transformation through a feed-forward neural network, the feature matrix of the current layer is input into the next-layer multi-head attention module. After superimposing the relative position score matrix, the feature matrix of the next layer is calculated until the final recognition result, that is, the predicted label, is output through a multi-layer perceptron (MLP Head).

[0107] Embodiment 2

[0108] This embodiment provides an experimental verification of the effect of the method described in Embodiment 1.

[0109] This experiment uses the FAST-Vehicle dataset collected by the Mini-SAR system independently developed by Nanjing University of Aeronautics and Astronautics. The collection scenario is the Nanjing Dajinshan National Defense Park, and the dataset contains 9 types of typical military targets (such as tanks, armored vehicles, and airplanes). As Figure 5 shown, in the figure, EO represents the use of visible light imaging technology, and RI represents the use of infrared radiation imaging technology. Figure 5 It covers the observation sequence of 360° azimuth at the pitch angle of 15° to 45°. The training set and the test set are divided according to the collection batches (time), as shown in Table 1. The original SAR image resolution is 64×64 pixels. To adapt to the input of the ViT (Vision Transformer) model, it is uniformly interpolated and adjusted to 224×224 pixels. Three-view images with equal intervals of azimuth are used for spatial stitching to generate a stitched image of 224×672 pixels, which not only retains the continuity of the target pose change but also effectively controls the computational complexity. The computing environment of the experiment is the NVIDIA GeForce RTX 2080Ti platform (equipped with 11GB video memory), and the software framework is based on PyTorch1.9. The hyperparameter configuration of the ViT model is as follows: the SGD optimizer (β = 0.9) is used, the initial learning rate is set to 0.05 and adjusted with the cosine annealing strategy. The regularization strategy includes a weight decay coefficient of 0.9 and a DropPath ratio of 0.3, and the batch size is set to 128.

[0110] Table 1 Training set and test set of the FAST-Vehicle dataset

[0111]

[0112] To verify the effectiveness of the method proposed in this application and compare the performance of different position encoding strategies in the multi-view SAR target recognition task, as shown in Table 2. The comparison objects include: (1) a single-view VIT model combined with learnable position encoding, that is, the standard VIT model, and this embodiment sets this mode as the baseline model; (2) a multi-view model combined with 4 different position encodings, namely absolute position encoding, relative position encoding, learnable position encoding, and the polar coordinate position encoding proposed in this application.

[0113] Table 2 Performance Comparison of Different Position Encoding Strategies

[0114]

[0115] As shown in Table 2, the baseline model (a single-view ViT model combined with learnable position encoding) achieved a test accuracy of 54.6% with the least number of parameters, constituting the performance benchmark. However, when multi-view features were introduced, although the number of parameters of each model increased significantly to more than 91M, the recognition accuracy of the multi-view model using absolute position encoding decreased instead of increasing, slightly inferior to the baseline model, only 53.4%. This indicates that the combination of simple early fusion and absolute position encoding may lead to performance degradation due to spatial confusion. Looking at other multi-view models, the correct rates of relative position encoding and ordinary learnable position encoding did not show obvious advantages, even lower than the baseline model. Only the polar coordinate position encoding model proposed in this application performed outstandingly, and the test accuracy was significantly improved to 56.8%, verifying its effectiveness and attributing it to its more accurate capture of the relative spatial relationship between multi-view images.

[0116] As Figure 6 shown, the confusion matrix of the method in this application on the test set is presented. Generally speaking, the diagonal values in the confusion matrix are significantly higher than the non-diagonal values, indicating that the method in this application has achieved high accuracy in the classification of various categories. For example, for the categories "54TH", "59AG", "63AT", and "T34-85", the classification accuracy rates are all over 80%. In addition, the non-diagonal values are generally low, indicating that the method in this application has a low degree of confusion between different categories and can effectively distinguish different categories. Figure 6 It intuitively verifies the effectiveness of the method in this application in improving the classification accuracy of multi-view images.

[0117] To more comprehensively evaluate the effectiveness of the method in this application, it is further compared with traditional image registration strategies. The traditional strategy relies on image registration preprocessing to eliminate perspective differences, and then extracts target features based on ViT by embedding absolute position encoding in image patches. The experimental results show that the test accuracy of this strategy is 54.2%, significantly lower than 56.8% of the method in this application.

[0118] Table 3 Performance Comparison with Traditional Image Registration Strategies

[0119]

[0120] The experimental results demonstrate the excellent performance of the method of this application in capturing multi - perspective spatial relationships, and its effectiveness significantly surpasses traditional image registration strategies. Although image registration can initially achieve perspective alignment, the processing inevitably causes image feature distortion, thereby affecting the recognition performance of the model. In contrast, the method proposed in this paper requires no image pre - processing and can directly obtain more robust and discriminative feature representations by learning features based on the original multi - perspective images, which reflects its superiority in the efficient utilization of multi - perspective information and proves the feasibility of replacing image registration.

[0121] To verify the performance advantages of the polar coordinate position encoding in the method proposed in this application, ablation experiments were designed to isolate and evaluate the independent contributions of the position encoding methods. In the experiments, the position score matrix was fixed as the spatial feature embedding method, and the performance differences between the polar coordinate position encoding proposed in this application and ordinary learnable position encoding were compared. The results in Table 4 show that the accuracy rate using polar coordinate position encoding is 56.8%, significantly better than 53.6% of ordinary learnable position encoding. This result proves that under the same spatial feature embedding framework, the polar coordinate position encoding proposed in this application can more effectively guide the model to learn the spatial information of multi - perspective images, thus significantly improving the classification performance.

[0122] Table 4 Comparison of ablation experiment results

[0123]

[0124] Experiments show that the method of this application has improved recognition accuracy on the FAST - Vehicle dataset, significantly better than traditional methods and single - perspective models. The comparative experiments verify the advantages of polar coordinate position encoding over image registration, and the ablation experiments confirm its effectiveness and the importance of appropriate perspective intervals. The method of this application realizes early feature fusion without registration, provides new ideas for the utilization of multi - perspective remote sensing image information, and demonstrates the potential to improve target recognition performance under complex conditions. In the future, adaptive angle interval adjustment strategies and more refined spatial feature embedding will be explored.

[0125] Embodiment 3

[0126] This embodiment provides a multi - perspective SAR target recognition system, including:

[0127] An image acquisition module, configured to acquire several SAR images with different perspectives and splice the several SAR images with different perspectives in the spatial dimension to obtain a spliced image;

[0128] An image division module, configured to divide each SAR image into a plurality of image patches and extract initial features of each image patch;

[0129] A polar coordinate encoding module, configured to determine a global pole point in the stitched image and establish a global polar coordinate system based on the global pole point, and calculate global polar coordinate position encodings of each image patch based on the global polar coordinate system;

[0130] A matrix calculation module, configured to calculate a relative position score matrix for quantifying the relative positions between image patches based on the global polar coordinate position encodings of each image patch;

[0131] An output module, configured to input the relative position score matrix and the initial features of each image patch into a pre-obtained Transformer model, construct a feature attention matrix of the image patches, superimpose the relative position score matrix and the feature attention matrix, and output a final recognition result.

[0132] Embodiment 4

[0133] This embodiment provides a multi-view SAR target recognition device, including a processor and a storage medium; the storage medium is used to store instructions; the processor is configured to operate according to the instructions to execute the steps of the multi-view SAR target recognition method described in Embodiment 1.

[0134] Embodiment 5

[0135] This embodiment provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the multi-view SAR target recognition method described in Embodiment 1 are implemented.

[0136] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0137] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, so that the instructions executed by the processors of the computer or other programmable data processing devices generate for implementing in the processFigure 1 a process or multiple processes and / or blocks Figure 1 a device for the functions specified in a block or multiple blocks.

[0138] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device that implements the functions in the process Figure 1 a process or multiple processes and / or blocks Figure 1 specified in a block or multiple blocks.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in the process Figure 1 a process or multiple processes and / or blocks Figure 1 specified in a block or multiple blocks.

[0140] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A multi-view SAR target recognition method, characterized in that Including: Obtain SAR images from several different perspectives, and splice the SAR images from several different perspectives in the spatial dimension to obtain a spliced image; Divide each SAR image into several image patches, and extract the initial features of each image patch; Determine the global poles in the spliced image and establish a global polar coordinate system based on the global poles, and calculate the global polar coordinate position encodings of each image patch based on the global polar coordinate system; Calculate a relative position score matrix for quantifying the relative positions between image patches according to the global polar coordinate position encodings of each image patch; Input the relative position score matrix and the initial features of each image patch into a pre-obtained Transformer model, construct a feature attention matrix of the image patches, superimpose the relative position score matrix and the feature attention matrix, and output the final recognition result.

2. The multi-view SAR target recognition method according to claim 1, characterized in that: Each perspective corresponds to one of the SAR images.

3. The multi-view SAR target recognition method according to claim 1, characterized in that: Dividing each of the SAR images into several image patches and extracting the initial features of each image patch includes: The several image patches are of the same size and non-overlapping; Extract the initial features of each image patch through a convolutional neural network model.

4. The multi-view SAR target recognition method according to claim 1, characterized in that: Determining the global poles in the spliced image and establishing a global polar coordinate system based on the global poles, and calculating the global polar coordinate position encodings of each image patch based on the global polar coordinate system includes: Taking the image center of any SAR image as the current local pole, a local polar coordinate system is established with the horizontal right as the polar axis, and the local polar coordinates of each image block are expressed as , where represents the distance from the current local pole of the i-th SAR image to the center point of the p-th image block, represents the distance from the current local pole of the i-th SAR image to the center point of the p-th image block and the angle with the polar axis; In the stitched image, taking the image center of any SAR image as the global pole, and establishing a global polar coordinate system with the direction pointing to the corresponding radar of the SAR image as the polar axis. In the global polar coordinate system, the distance offset of the image center of the i-th SAR image relative to the global pole is expressed as , and the viewing angle offset of the corresponding radar of the i-th SAR image relative to the global pole is expressed as ; Combining the local polar coordinate system and the global polar coordinate system, the global polar coordinate position encoding of the p-th image patch of the i-th SAR image is expressed as: ； Among them, represents the global polar coordinate position encoding of the p-th image patch in the i-th SAR image.

5. The multi-view SAR target recognition method according to claim 4, characterized in that Calculating a relative position score matrix for quantifying the relative positions between image patches according to the global polar coordinate position encodings of each image patch includes: The global polar coordinate position encodings of any two image patches are respectively expressed as: ； ； Among them, represents the global polar coordinate position encoding of the q-th image patch in the j-th SAR image, represents the distance from the current local pole of the j-th SAR image to the center point of the q-th image patch, represents the distance from the current local pole of the j-th SAR image to the center point of the q-th image patch and the angle with the polar axis, represents the distance offset of the image center of the j-th SAR image relative to the global pole, represents the viewing angle offset of the corresponding radar of the j-th SAR image relative to the global pole; Calculate the distance between two image patches , expressed as: ； When the distance between two image blocks is expressed as: ； The distance matrix of all image patches is expressed as: ； Among them, D represents the distance matrix of all image patches, M represents the total number of image patches, represents the distance between the p-th image patch and the q-th image patch, ; Linearly compress each element in the distance matrix D, which is expressed as: ； Among them, represents the distance between two linearly compressed image blocks, represents the maximum value of the elements in the distance matrix D; The relative position score matrix is expressed as: ； Among them, represents the relative position score matrix, represents the compressed distance matrix, , represents the compressed distance between the p-th image block and the q-th image block, represents the weight parameter, represents the bias parameter.

6. The multi-view SAR target recognition method according to claim 1, wherein Inputting the relative position score matrix and the initial features of each image patch into a pre-obtained Transformer model, constructing a feature attention matrix of the image patches, superimposing the relative position score matrix and the feature attention matrix, and outputting the final recognition result includes: The Transformer model includes multiple feature extraction layers, and each feature extraction layer includes a multi-head attention module; The initial features of each image patch are input into the multi-head attention module of the first layer to obtain a feature attention matrix, which is expressed as: ； Among them, represents the feature attention matrix, Q represents the query matrix, K represents the key matrix, and T represents the matrix transpose, represents the vector dimension; In the multi-head attention module of the first layer, superimpose the relative position score matrix and the feature attention matrix, and calculate the first layer feature matrix, which is expressed as: ； Among them, represents the first-layer feature matrix, represents the relative position score matrix, V represents the value matrix, represents the softmax operation; Input the feature matrix of the current layer into the multi-head attention module of the next layer, superimpose the relative position score matrix, and calculate the feature matrix of the next layer until the final recognition result is output.

7. A multi-view SAR target recognition system, characterized in that, Including: An image acquisition module, configured to obtain SAR images from several different perspectives, and splice the SAR images from several different perspectives in the spatial dimension to obtain a spliced image; An image division module, configured to divide each SAR image into several image patches, and extract the initial features of each image patch; A polar coordinate encoding module, which is used to determine the global poles in the spliced image, establish a global polar coordinate system based on the global poles, and calculate the global polar coordinate position encoding of each image block based on the global polar coordinate system; A matrix calculation module, which is used to calculate a relative position score matrix for quantifying the relative positions between image blocks according to the global polar coordinate position encoding of each image block; An output module, which is used to input the relative position score matrix and the initial features of each image block into a pre-acquired Transformer model, construct a feature attention matrix of the image blocks, superimpose the relative position score matrix and the feature attention matrix, and output a final recognition result.

8. A multi-view SAR target recognition device, characterized in that It includes a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the multi-view SAR target recognition method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the multi-view SAR target recognition method according to any one of claims 1 to 6.

Images