SAR image target recognition method, system, electronic equipment and medium

By converting optical SAR images into tensors and performing high-order singular value decomposition to obtain the basis matrix and performing low-rank approximation metric, a deep neural network is trained. This solves the problem of uninterpretable features in target recognition of optical SAR images by deep networks, thus improving the interpretability and accuracy of recognition.

CN117253098BActive Publication Date: 2024-11-26PINGHU SPACE PERCEPTION LAB TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311186955.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-14
Publication Date
2024-11-26
Estimated Expiration
2043-09-14

AI Technical Summary

Technical Problem

The features extracted by deep networks during target recognition in optical SAR images cannot be interpreted, resulting in limited recognition performance.

Method used

The optical SAR image is converted into a tensor, and a high-order singular value decomposition is performed to obtain the basis matrix. The distance metric of the basis matrix is ​​calculated by low-rank approximation. Similar basis matrices are placed in the same subspace and a deep neural network is trained for target recognition.

Benefits of technology

It improves the interpretability and accuracy of target recognition in optical SAR images, solves the problem of the inability to interpret features from deep networks, and achieves higher-level feature extraction and classification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117253098B_ABST
    Figure CN117253098B_ABST
Patent Text Reader

Abstract

The present invention relates to a SAR image target recognition method, system, electronic device and medium. The method comprises acquiring multiple optical SAR images of different categories and converting each optical SAR image into a corresponding tensor; performing high-order singular value decomposition on each tensor to determine the basis matrix corresponding to each tensor; calculating the distance metric between any two basis matrices by low-rank approximation, and if the distance metric is less than a preset value, placing the two basis matrices in the same preset subspace; using the basis matrices in the same preset subspace as a basis matrix set, and using the basis matrix sets corresponding to different preset subspaces as a training set to train a deep neural network to obtain a recognition model; and performing target recognition on the SAR image by using the recognition model. The problem that the features extracted by the deep network cannot be explained in the target recognition process of the optical SAR image is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Background Art

[0002] Synthetic Aperture Radar (SAR) is an active earth observation system based on a multi-motion platform. Due to the penetrating power of electromagnetic waves, SAR is not limited by light and climate conditions and has many characteristics such as all-day, all-weather, high resolution and wide bandwidth. After decades of development, SAR imaging technology has gradually matured, and the data acquisition capability of high-resolution SAR images has been significantly improved, laying the foundation for the intelligent application of SAR images. SAR target recognition is to classify images in feature space, so extracting effective features from images is the key to improving recognition performance.

[0003] Many researchers have begun to try to use convolutional neural networks to deal with SAR image target recognition problems and have achieved ideal results. Although CNN-based methods can easily obtain satisfactory recognition results, the features extracted by deep networks are often unexplainable, resulting in unexplainable problems that need to be solved. Summary of the invention

[0004] In order to overcome the problem that features extracted by deep networks cannot be interpreted during target recognition in optical SAR images, the present invention provides a SAR image target recognition method, system, electronic device and medium.

[0005] In a first aspect, in order to solve the above technical problems, the present invention provides a SAR image target recognition method, comprising:

[0006] Obtain multiple optical SAR images of different categories and convert each optical SAR image into a corresponding tensor;

[0007] Perform high-order singular value decomposition on each tensor to determine the basis matrix corresponding to each tensor;

[0008] The distance metric between any two basis matrices is calculated by low-rank approximation. If the distance metric is less than a preset value, the two basis matrices are placed in the same preset subspace, where one preset subspace corresponds to one category of optical SAR images.

[0009] The basis matrices in the same preset subspace are used as a basis matrix set, and the basis matrix sets corresponding to different preset subspaces are used as training sets to train the deep neural network to obtain a recognition model;

[0010] Target recognition is performed on SAR images through recognition models.

[0011] In a second aspect, the present invention provides a SAR image target recognition system, comprising:

[0012] A tensor acquisition module, used to acquire multiple optical SAR images of different categories and convert each optical SAR image into a corresponding tensor;

[0013] A basis matrix determination module is used to perform high-order singular value decomposition on each tensor to determine the basis matrix corresponding to each tensor;

[0014] A distance metric module is used to calculate the distance metric between any two basis matrices through low-rank approximation. If the distance metric is less than a preset value, the two basis matrices are placed in the same preset subspace, where one preset subspace corresponds to one category of optical SAR images.

[0015] A training module is used to train a deep neural network by taking the basis matrices in the same preset subspace as a basis matrix set and taking the basis matrix sets corresponding to different preset subspaces as training sets to obtain a recognition model;

[0016] The target recognition module is used to perform target recognition on SAR images through the recognition model.

[0017] In a third aspect, the present invention further provides an electronic device, comprising a memory, a processor, and a program stored in the memory and running on the processor, wherein when the processor executes the program, the steps of a SAR image target recognition method as described above are implemented.

[0018] In a fourth aspect, the present invention further provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes the steps of a SAR image target recognition method.

[0019] The beneficial effects of the present invention are as follows: an optical SAR image is converted into a tensor, and then a high-order singular value decomposition is performed on the tensor to obtain a corresponding basis matrix. The high-order singular value decomposition of the tensor can obtain useful information expressed at a higher level in the optical SAR image, which is convenient for a subsequent neural network to effectively extract features in the optical SAR image, and then the distance metric between any two basis matrices is calculated by low-rank approximation, the optical SAR image is classified, and the deep neural network is trained after the training set is classified to obtain a recognition model, and finally the SAR image is used to recognize targets through the recognition model. Since the recognition model extracts the basis matrix after the high-order singular value decomposition, it has useful information expressed at a higher level, and solves the problem that the features extracted by the deep network cannot be explained in the target recognition process of the optical SAR image. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0021] Figure 1A schematic diagram of a flow chart of a SAR image target recognition method according to an embodiment of the present invention;

[0022] Figure 2 Schematic diagram of tensor approximation;

[0023] Figure 3 Schematic diagram of tensor decomposition;

[0024] Figure 4 is the geometric relationship between the target position and the antenna beam;

[0025] Figure 5 Identify the results for the confusion matrix;

[0026] Figure 6 The figure is a schematic diagram of the structure of a SAR image target recognition system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following examples are provided to further explain and supplement the present invention and do not constitute any limitation to the present invention.

[0028] A SAR image target recognition method, system, electronic device and medium according to an embodiment of the present invention are described below with reference to the accompanying drawings.

[0029] like Figure 1 As shown, a SAR image target recognition method according to an embodiment of the present invention comprises the following steps:

[0030] S1. Acquire multiple optical SAR images of different categories and convert each optical SAR image into a corresponding tensor.

[0031] A first-order tensor is an object with an index. The dimension of a tensor is called a mode. Common vectors and matrices can be regarded as first-order tensors and second-order tensors, respectively. However, there are three color channels, R, G, and B, in optical SAR images. Therefore, a third-order tensor is needed to represent optical SAR images.

[0032] The third-order tensor can be expressed as Among them, A represents a third-order tensor, I, J, and K are all positive numbers, which determine the number of elements in each order in the third-order tensor.

[0033] In addition, since each order in the third-order tensor contains multiple elements, the tensor can be matrixed. The modulus n matrixing of the tensor K is a common operation, where the modulus n is the fiber arrangement as the column of the matrix, denoted as K (n) , and then by specifying The matrixization of the third-order tensor A is defined by the position of the elements in the matrix, as follows:

[0034]

[0035] Among them, A (n) The column vector of is the modulo n fiber of A, a ijk represents any element in a third-order tensor, They represent the elements of the lth, Jth and Kth columns of the mode n fiber respectively, v represents a specific value, i, j, k represent index values ​​respectively, and the three index values ​​determine the position of an element in the third-order tensor.

[0036] S2. Perform high-order singular value decomposition on each tensor to determine the basis matrix corresponding to each tensor.

[0037] The high-order singular value decomposition of tensors involves the multiplication of tensors and matrices. Therefore, for any tensor and any matrix The n-modular tensor and matrix product of A and F is represented by A× n F, calculated as follows:

[0038]

[0039] Among them, i1,…i N Represents the elements in the tensor, j n ,i n Represents an element in a matrix.

[0040] Given a third-order tensor matrix matrix and matrix The product of a third-order tensor and a matrix satisfies:

[0041]

[0042] The above describes the product of a third-order tensor and a matrix. Therefore, Figure 2 As shown, the high-order singular value decomposition of the third-order tensor can be expressed as:

[0043]

[0044] Among them, A represents a third-order tensor, and is the basis matrix obtained by decomposing the third-order tensor, J is a real tensor with the same dimension as A, and satisfies the following conditions:

[0045] (1) (Full orthogonality) Any two different slices (a slice is a tensor determined by three index values. After a dimension value is determined, the remaining two dimensions represent a slice) fixed in the same pattern are orthogonal;

[0046]

[0047] J(v,:,:) and J(λ,:,:) represent the slices corresponding to the remaining two dimensions after the first dimension is determined. J(:,v,:) and J(:,λ,:) represent the slices corresponding to the remaining two dimensions after the second dimension is determined. J(:,:,v) and J(:,:,λ) represent the slices corresponding to the remaining two dimensions after the third dimension is determined. v and λ respectively represent the dimensions corresponding to the slices.

[0048] (2) (Order) The norms of the slices along each mode are ordered. For example, for the first mode, we have:

[0049] ||J(1,:,:)||≥||J(2,:,:)||≥…≥0

[0050] The norm in the formula is actually a matrix tensor Singular values ​​of, third-order tensor The high-order singular value decomposition of can be calculated by calculating the orthogonal matrices U, V and W as A (n) (n-module tensor) is done as follows:

[0051] (a) Calculate three orthogonal matrices:

[0052]

[0053] (b) Calculate the core tensor:

[0054] J=A×1U T ×2V T ×3W T

[0055] Based on the above, the tensor approximation can be expressed as:

[0056]

[0057] in, and

[0058] Since the norm of the omitted part of the core tensor J is small, the approximation error is small. Therefore, the high-order singular value decomposition of the third-order tensor can be rewritten as follows:

[0059]

[0060] Among them, Figure 3 As shown, the square represents A v , the title line above it is w which represents the outer product of the third mode v The vector is the linear combination coefficient, J(:,:,v)×1U×2V represents the product of the remaining two slices and the third dimension after the third dimension is determined.

[0061] S3. Calculate the distance metric between any two basis matrices by low-rank approximation. If the distance metric is less than a preset value, put the two basis matrices into the same preset subspace, where one preset subspace corresponds to a category of optical SAR images.

[0062] S4. Taking the basis matrix in the same preset subspace as a basis matrix set, and taking the basis matrix sets corresponding to different preset subspaces as training sets to train the deep neural network, so as to obtain a recognition model.

[0063] S5. Perform target recognition on SAR images through recognition models.

[0064] In this embodiment, the optical SAR image is converted into a tensor, and then the tensor is subjected to high-order singular value decomposition to obtain a corresponding basis matrix. The high-order singular value decomposition of the tensor can obtain useful information expressed at a higher level in the optical SAR image, which is convenient for the subsequent neural network to effectively extract features in the optical SAR image. The distance metric between any two basis matrices is calculated by low-rank approximation, the optical SAR image is classified, and the deep neural network is trained after the training set is classified to obtain a recognition model. Finally, the SAR image is used to recognize targets. Since the recognition model extracts the basis matrix after high-order singular value decomposition, it has useful information expressed at a higher level, which solves the problem that the features extracted by the deep network cannot be explained in the target recognition process of the optical SAR image.

[0065] Optionally, perform high-order singular value decomposition on each tensor to determine the basis matrix corresponding to each tensor. The formula is as follows:

[0066]

[0067] Among them, A SAR represents a tensor, The basis matrix corresponding to the v-th tensor is, represents the linear combination coefficients, K represents the first K columns of the tensor, and ×3 represents the modulo 3 product.

[0068] By performing high-order singular value decomposition on the third-order tensor, the corresponding basis matrix is ​​obtained. The basis matrix has useful information expressed at a higher level, which facilitates the extraction of useful feature information during subsequent deep neural network training.

[0069] Optionally, the distance metric between any two basis matrices is calculated by low-rank approximation, as follows:

[0070]

[0071] Among them, F represents the distance metric, U k and V kare the first k columns of U and V respectively, U and V represent any two basis matrices, ∑ k =diag(σ1,…,σ k ), diag() represents the diagonalization operator, σ1, L, σ k Each represents k singular values.

[0072] Through low-rank approximation, the distance metric between any two basis matrices can be obtained, so that two three-dimensional tensors with a distance metric less than a preset value can be projected into the same preset subspace, thereby classifying different optical SAR images.

[0073] Optionally, the method also includes testing the recognition model, including: obtaining multiple test images of different categories, obtaining the recognition result corresponding to each test image through the recognition model; calculating the square norm of each test image and the corresponding recognition result; and evaluating the recognition model according to each square norm.

[0074] By testing the recognition model, the performance of the recognition model is analyzed.

[0075] Optionally, the square norm of each test image and the corresponding recognition result is calculated as follows:

[0076]

[0077] Among them, R(μ) represents the square norm, D represents the recognition result corresponding to the test image, represents the normalized value of the tensor corresponding to the vth test image, k represents the first k columns of the vth tensor, and μ represents.

[0078] The derivation formula of the square norm is as follows:

[0079] For unknown optical SAR images, consider the minimization problem as follows:

[0080]

[0081] in, represents an unknown optical SAR image, express The corresponding tensor, D represents the recognition result corresponding to the unknown optical SAR image, and K represents the first k columns of the vth tensor.

[0082] The solution to the above minimization problem is as follows:

[0083]

[0084] Inserting the result of formula (2) into formula (1) yields:

[0085]

[0086] Optionally, multiple optical SAR images of different categories are obtained, and each optical SAR image is converted into a corresponding tensor, including:

[0087] Get the target echo corresponding to the nth pulse and the mth range unit,

[0088]

[0089] Among them, s r (m,n) represents the target echo, σ p represents the target scattering coefficient, x0 is the starting point of the radar platform, T PRT represents the pulse repetition time, v is the radar platform speed, x T is the azimuth coordinate, c is the speed of light, T p Indicates the pulse width of the transmitted signal, f s is the sampling rate in the distance dimension, k r is the frequency modulation slope of the linear FM signal, R n is the relative distance between the range unit and the antenna phase center;

[0090] According to the working parameters of the radar platform and each target echo, the radar data corresponding to each target echo is determined by traversing and;

[0091] Perform a two-dimensional Fourier transform on each radar data to obtain the corresponding optical SAR image;

[0092] Convert an optical SAR image to a corresponding tensor.

[0093] The conversion of optical SAR images into corresponding tensors is a prior art and will not be described in detail.

[0094] The derivation process of the target echo is as follows:

[0095] like Figure 4 As shown in the figure, it is the geometric relationship between the target position and the antenna wave speed. From this geometric relationship, it can be seen that if there is a point target located at P(x T ,R0), when the radar transmits the nth pulse, the relative distance between the target and the antenna phase center can be expressed as:

[0096]

[0097] Among them, R0 is the distance coordinate, x T is the azimuth coordinate, x0 is the starting point of the radar platform, β is the 3dB beam width, v is the platform speed, T PRT Indicates the pulse repetition time.

[0098] After down-conversion and IQ demodulation, the radar echo corresponding to the nth transmitted pulse can be expressed as:

[0099]

[0100] Among them, s r (m,n) represents radar echo, σ p represents the target scattering coefficient, x0 is the starting point of the radar platform, T PRT represents the pulse repetition time, v is the radar platform speed, x T is the azimuth coordinate, c is the speed of light, T p represents the pulse width of the transmitted signal, k r is the frequency modulation slope of the linear FM signal, R n is the relative distance between the distance unit and the antenna phase center.

[0101] Assume that the sampling rate of the distance dimension is f s , discretize the radar echo in formula (4) to get the target echo.

[0102] Optionally, in order to better illustrate this embodiment, a specific example is used for illustration:

[0103] In this embodiment, when K=60 in the first K columns of the tensor, Figure 5 The confusion matrix of the recognition results is shown. Figure 5 It can be seen that all the diagonal elements are close to 1, which means that the recognition results of each type of object are excellent.

[0104] like Figure 6 As shown, the present invention provides a SAR image target recognition system, comprising:

[0105] A tensor acquisition module 101 is used to acquire multiple optical SAR images of different categories and convert each optical SAR image into a corresponding tensor;

[0106] A basis matrix determination module 102 is used to perform high-order singular value decomposition on each tensor to determine a basis matrix corresponding to each tensor;

[0107] A distance metric module 103 is used to calculate the distance metric between any two basis matrices by low-rank approximation. If the distance metric is less than a preset value, the two basis matrices are placed in the same preset subspace, where one preset subspace corresponds to one category of optical SAR images.

[0108] A training module 104 is used to train a deep neural network using base matrices in the same preset subspace as a base matrix set and base matrix sets corresponding to different preset subspaces as training sets to obtain a recognition model;

[0109] The target recognition module 105 is used to perform target recognition on the SAR image through a recognition model.

[0110] Optionally, the basis matrix determination module 102 is specifically used for:

[0111] Perform high-order singular value decomposition on each tensor to determine the basis matrix corresponding to each tensor. The formula is as follows:

[0112]

[0113] Among them, A SAR represents a tensor, The basis matrix corresponding to the vth tensor is, represents the linear combination coefficient, K represents the first K columns of the tensor, and ×3 represents.

[0114] Optionally, the distance measurement module 103 is specifically used for:

[0115] The distance metric between any two basis matrices is calculated by low-rank approximation, as follows:

[0116]

[0117] Among them, F represents the distance metric, U k and V k are the first k columns of U and V, respectively. U and V represent any two basis matrices, Σ k =diag(σ1,…,σ k ), diag() represents the diagonalization operator, σ1, L, σ k Each represents k singular values.

[0118] Optionally, the system further includes a testing module, specifically configured to:

[0119] Obtain multiple test images of different categories, and obtain the recognition result corresponding to each test image through the recognition model;

[0120] Calculate the square norm of each test image and the corresponding recognition result;

[0121] The recognition models were evaluated according to the respective squared norms.

[0122] Optionally, a test module is specifically used to:

[0123] Calculate the square norm of each test image and the corresponding recognition result. The formula is as follows:

[0124]

[0125] μ=0,1,…,9.

[0126] Among them, R(μ) represents the square norm, D represents the recognition result corresponding to the test image, represents the normalized value of the tensor corresponding to the vth test image, k represents the first k columns of the vth tensor, and μ represents.

[0127] Optionally, the tensor acquisition module 101 is specifically used for:

[0128] Obtain multiple optical SAR images of different categories and convert each optical SAR image into a corresponding tensor, including:

[0129] Get the target echo corresponding to the nth pulse and the mth range unit,

[0130]

[0131] Among them, s r (m,n) represents the target echo, σ p represents the target scattering coefficient, x0 is the starting point of the radar platform, T PRT represents the pulse repetition time, v is the radar platform speed, x T is the azimuth coordinate, c is the speed of light, T p Indicates the pulse width of the transmitted signal, f s is the sampling rate in the distance dimension, k r is the frequency modulation slope of the linear FM signal, R n is the relative distance between the range unit and the antenna phase center;

[0132] According to the working parameters of the radar platform and each target echo, the radar data corresponding to each target echo is determined by traversing and;

[0133] Perform a two-dimensional Fourier transform on each radar data to obtain the corresponding optical SAR image;

[0134] Convert an optical SAR image to a corresponding tensor.

[0135] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, some or all steps of the above-mentioned SAR image target recognition method are implemented.

[0136] Among them, the electronic device can be a computer, and correspondingly, its program is computer software. The above-mentioned parameters and steps in an electronic device of the present invention can refer to the parameters and steps in an embodiment of a SAR image target recognition method above, and will not be repeated here.

[0137] Those skilled in the art know that the present invention can be implemented as a system, method or computer program product. Therefore, the present disclosure can be specifically implemented in the following forms, namely: it can be complete hardware, it can be complete software (including firmware, resident software, microcode, etc.), or it can be a combination of hardware and software, generally referred to as "circuit", "module" or "system" herein. In addition, in some embodiments, the present invention can also be implemented in the form of a computer program product in one or more computer-readable media, and the computer-readable medium contains computer-readable program code. Computer-readable storage media can be, for example, but not limited to - electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or devices, or any combination of the above.

[0138] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0139] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A SAR image target recognition method, It is characterized in that include: Obtain multiple optical SAR images of different categories and convert each optical SAR image into a corresponding tensor; Perform high-order singular value decomposition on each tensor to determine the basis matrix corresponding to each tensor; Calculating a distance metric between any two basis matrices by low-rank approximation, and if the distance metric is less than a preset value, placing the two basis matrices into the same preset subspace, wherein the preset subspace corresponds to a category of an optical SAR image; Taking the basis matrix in the same preset subspace as a basis matrix set, and taking the basis matrix sets corresponding to different preset subspaces as training sets to train a deep neural network, to obtain a recognition model; Target recognition is performed on SAR images through recognition models; The tensor is subjected to high-order singular value decomposition to determine the basis matrix corresponding to each tensor. The formula is as follows: in, represents a tensor, The basis matrix corresponding to the vth tensor is, represents the linear combination coefficient, K represents the first K columns of the tensor, × 3 represents modulo 3 product; The distance metric between any two basis matrices is calculated by low-rank approximation, as follows: Among them, F represents the distance metric, U k and V k are the first k columns of U and V, respectively. U and V represent any two basis matrices, Σ k =diag(σ 1 ,…,σ k ), diag() represents the diagonalization operator, σ 1 , …, σ k represents k singular values.

2. The method according to claim 1, It is characterized in that The method also includes testing of the recognition model, including: Obtain multiple test images of different categories, and obtain the recognition result corresponding to each test image through the recognition model; Calculate the square norm of each test image and the corresponding recognition result; The recognition model is evaluated according to each of the squared norms.

3. The method according to claim 2, It is characterized in that Calculate the square norm of each test image and the corresponding recognition result. The formula is as follows: Among them, R(μ) represents the square norm, D represents the recognition result corresponding to the test image, represents the normalized value of the tensor corresponding to the vth test image, k represents the first k columns of the vth tensor, and μ represents the sample category.

4. The method according to any one of claims 1 to 3, It is characterized in that Obtain multiple optical SAR images of different categories and convert each optical SAR image into a corresponding tensor, including: Get the target echo corresponding to the nth pulse and the mth range unit, Among them, s r (m,n) represents the target echo, σ p represents the target scattering coefficient, x 0 is the starting point of the radar platform, T PRT represents the pulse repetition time, v is the radar platform speed, x T is the azimuth coordinate, c is the speed of light, T p Indicates the pulse width of the transmitted signal, f s is the sampling rate in the distance dimension, k r is the frequency modulation slope of the linear FM signal, R n is the relative distance between the range unit and the antenna phase center; According to the working parameters of the radar platform and each target echo, the radar data corresponding to each target echo is determined through two-dimensional Fourier transform; Perform a two-dimensional Fourier transform on each radar data to obtain the corresponding optical SAR image; The optical SAR image is converted into a corresponding tensor.

5. A SAR image target recognition system, It is characterized in that include: A tensor acquisition module, used to acquire multiple optical SAR images of different categories and convert each optical SAR image into a corresponding tensor; A basis matrix determination module is used to perform high-order singular value decomposition on each tensor to determine the basis matrix corresponding to each tensor; A distance metric module, used for calculating the distance metric between any two basis matrices by low-rank approximation, and if the distance metric is less than a preset value, the two basis matrices are placed in the same preset subspace, where the preset subspace corresponds to a category of an optical SAR image; A training module, used to use the basis matrix in the same preset subspace as a basis matrix set, and use the basis matrix sets corresponding to different preset subspaces as training sets to train a deep neural network to obtain a recognition model; A target recognition module is used to perform target recognition on SAR images through a recognition model; The basis matrix determination module is specifically used for: The tensor is subjected to high-order singular value decomposition to determine the basis matrix corresponding to each tensor. The formula is as follows: in, represents a tensor, The basis matrix corresponding to the vth tensor is, represents the linear combination coefficient, K represents the first K columns of the tensor, × 3 express; The distance measurement module is specifically used for: The distance metric between any two basis matrices is calculated by low-rank approximation, as follows: Among them, F represents the distance metric, U k and V k are the first k columns of U and V, respectively. U and V represent any two basis matrices, Σ k =diag(σ 1 ,…,σ k ), diag() represents the diagonalization operator, σ 1 , L, σ k Each represents k singular values.

6. An electronic device comprising a memory, a processor, and a program stored in the memory and running on the processor, It is characterized in that When the processor executes the program, the steps of the SAR image target recognition method according to any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the steps of a SAR image target recognition method according to any one of claims 1 to 4.