SAR Image Target Recognition Method Based on Generalized Principal Component Analysis and GAN

Abstract

The present invention proposes a SAR image target recognition method based on generalized principal component analysis and the generative adversarial network GAN, comprising the following steps: (1) obtaining a training sample set and a test sample set; (2) performing generalized principal component analysis on the training sample set; (3) constructing a generative adversarial network model; (4) performing iterative training on the generative adversarial network model; (5) obtaining the SAR image target recognition result. The present invention takes the optimal projection matrix obtained by performing generalized principal component analysis on the training sample set as a part of the weight parameters of the generative network G and the discriminative network D, and then obtains an optimal generative adversarial network model through iterative training, so that the generative network G can generate the azimuth target pseudo-images missing in the training sample set, and the discriminative network D can obtain more complete target azimuth information and structural information, solving the problem that the prior art does not effectively utilize the target azimuth information and improving the correct rate of SAR image target recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and relates to a method for SAR image target recognition. Specifically, it relates to a method for SAR image target recognition based on generalized principal component analysis and generative adversarial network GAN, which can be used for target detection tasks under many platforms such as airborne and shipborne platforms. Background Art

[0002] Synthetic Aperture Radar (SAR), as an active observation system, can observe surface features all day and all weather, and has advantages such as strong penetration ability and resolution not affected by distance. SAR image target recognition is to process SAR images to achieve classification and discrimination of targets of interest, which is an important extension of radar detection function and has become a research hotspot in the current fields of SAR image interpretation and pattern recognition.

[0003] In recent years, with the rapid development of deep learning technology and large-scale open-source data sets, the ability of pattern recognition and classification has been greatly enhanced, bringing new ideas to the field of SAR image target recognition. Its deep network structure can automatically obtain target features when the number of target samples is sufficient, greatly improving the accuracy of SAR image target recognition. However, when recognizing SAR image targets with few target samples, the training process of deep learning technology has an underfitting problem. As an important part of the deep learning field, the generative adversarial network has a function of expanding the sample set with its mutually adversarial generator and discriminator network structures, which can solve the above underfitting problem. However, the generative adversarial network does not improve the azimuth completeness of the training samples, resulting in a still low accuracy of SAR image target recognition. Therefore, how to improve the accuracy of target recognition when the number of SAR image target samples is small is an urgent problem to be solved.

[0004] In order to ensure that the training process does not suffer from underfitting when the target sample size is small and at the same time improve the target recognition accuracy rate, the R & D personnel proposed a small-sample SAR image target recognition method that uses an improved generative adversarial network combined with a convolutional neural network. For example, in the patent application with the application publication number CN 113298007 A and the name "A Small-Sample SAR Image Target Recognition Method", this method first established a generative adversarial network based on the characteristics of SAR image targets, and then introduced two types of objective parameters to evaluate the generated images, screened the images with higher evaluations to construct an expanded dataset. Finally, a SAR image classification model based on CNN was used to perform target recognition on the expanded dataset, which can improve the target recognition accuracy rate in some specific cases. However, this method directly uses the training samples to expand the samples of the generative adversarial network, only utilizes partial target structure information, and does not effectively utilize the target azimuth information, resulting in incomplete target azimuth in the expanded dataset and limited improvement in the target recognition accuracy rate. Moreover, using the convolutional neural network to perform target recognition on the expanded dataset still cannot obtain a complete description of the target azimuth information, ultimately leading to limited improvement in the target recognition accuracy rate. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects existing in the above-mentioned prior art, and propose a SAR image target recognition method based on generalized principal component analysis and generative adversarial network to solve the technical problem of low target recognition accuracy rate when the SAR image target sample size is small in the prior art.

[0006] To achieve the above purpose, the technical solution adopted by the present invention includes the following steps:

[0007] (1) Obtain a training sample set and a test sample set:

[0008] (1a) Obtain P SAR images T1 = {T 11 , T 12 , …, T 1p , …, T 1P} with ξ target categories and target labels when the synthetic aperture radar (SAR) has a working pitch angle of α, and Q SAR images T2 = {T 21 , T 22 , …, T 2q , …, T 2Q} when the working pitch angle is β, where ξ ≥ 2, T 1p represents the p-th SAR image with a pitch angle of α and containing R α targets, R α ≥ 1, P ≥ 20ξR α , T 2q represents the p-th SAR image with a pitch angle of β and containing R βThe qth SAR image of a target, R β ≥1, Q≥10ξ / R β ;

[0009] (1b) A cropping box of size m×n is used to crop each SAR image T 1p Included R α targets, and each SAR image T 2p Included R β Each target is center cropped and each cropped image block is normalized. Then the U normalized image blocks corresponding to T1 are normalized. trn ={I1,I2,…,I u ,…I U} form a training sample set, and the V normalized image blocks I corresponding to T2 test ={I1,I2,…,I v ,…I V} form a test sample set, where m≥100, n≥100, I u represents the u-th training sample, I v represents the vth test sample;

[0010] (2) Perform generalized principal component analysis on the training sample set:

[0011] (2a) Use a sliding window of size k1×k2 to train the sample set I trn ={I1,I2,,I u ,…I U Each sample I in u Perform single-step sliding window value acquisition to obtain I u The corresponding size is Image sub-block matrix X u , and I trn The corresponding U image sub-block matrices are cascaded to obtain a size of The joint matrix X, where k1≥3, k2≥3, Indicates rounding up;

[0012] (2b) Initialize the identity matrix Λ of size L×L L , the projection matrix W of size (k1×k2)×L is the unknown matrix, and W T W=Λ L , the generalized principal component projection reconstruction function is And perform S iterations to solve the undetermined matrix W in F(X,W) to obtain the optimal projection matrix set Among them, L≥1, S≥2, W s *The \(s\)-th optimal projection matrix of size \(k1\times k2\times L\), \(\mu\) represents the element average of the joint matrix \(X\), \(W\) T represents the transpose operation of \(W\), represents the operation of taking the two-norm;

[0013] (3) Construct the generative adversarial network model \(O\):

[0014] (3a) Construct the structure of the generative adversarial network model \(O\):

[0015] Construct a generative adversarial network model \(O\) including a sequentially connected generative network \(G\) and a discriminative network \(D\); where the generative network \(G\) adopts a two-dimensional transposed convolutional neural network composed of a fully connected layer, \(S\) generative basic blocks, and a tangent activation function layer connected in series, and the generative basic block includes a two-dimensional transposed convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected in series; the discriminative network \(D\) adopts a two-dimensional convolutional neural network composed of \(S\) discriminative basic blocks, a fully connected layer, and a Softmax classification output layer connected in series, and the discriminative basic block includes a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer connected in series;

[0016] (3b) Define the loss function \(L\) of the generative network \(G\) G and the loss function \(L\) of the discriminative network \(D\) D :

[0017] \(L\) G =E_f[\log(1 - D(G(z)))]

[0018] \(L\) D =E_f[\log D(I u )]+E_f[\log(1 - D(G(z)))]

[0019] where \(z\) represents a random sequence, \(G(\cdot)\) represents the transposed sampling function of the generative network \(G\), \(D(\cdot)\) represents the class prediction function of the discriminative network \(D\), and \(E_f[\cdot]\) represents the expectation function;

[0020] (4) Iteratively train the generative adversarial network model \(O\):

[0021] (4a) Initialize the weight parameters of the generative network \(G\) and the discriminative network \(D\) as \(\{W\) * ,\(\theta\) G \}\), \(\{W\) * ,\(\theta\) D \}\), the number of iterations is \(e\), the maximum number of iterations is \(E\), and let \(e = 0\), and at the same time fix \(W\) * not to participate in the iteration, where \(E\geq200\), \(\theta\) G represents the parameter set of the bias of the two-dimensional transposed convolutional layer and the weights and biases of the fully connected layer in \(G\), \(\theta\) DThe parameter set representing the bias of the two-dimensional convolutional layer and the weights and biases of the fully connected layer in D;

[0022] (4b) Use the random sequence z and the training sample set I trn As the input of the generative adversarial network model O, the generative network G performs transposed sampling on the random sequence z to obtain the pseudo-sample G e (z); The discriminative network D performs class prediction on the pseudo-sample G e (z) output by G, and each training sample I u to obtain the prediction vectors D e (z), I u for the ξ-dimensional classes; e (G e (z)), D e (I u );

[0023] (4c) Adopt the loss function L G of the generative network G, and calculate the loss value of G through the ξ-dimensional class prediction vector D e (z) of G e (G e (z)) Adopt the loss function L D of the discriminative network D, and calculate the loss value of D through the ξ-dimensional class prediction vector D e (z) of G e (G e (z)) and the ξ-dimensional class prediction vector D u of each training sample I e (I u );

[0024] (4d) Use the gradient descent method, and update the parameters θ of the generative network G through the partial derivative of the parameter , and update the parameters θ G of the discriminative network D through the partial derivative of the parameter ; to obtain the generative adversarial network model O D with the weight parameters of G as and the weight parameters of D as ; e ;

[0025] (4e) Judge whether e≥E holds. If so, obtain the trained generative adversarial network model O * , otherwise, set e = e + 1, O e = O, and execute step (4b);

[0026] (5) Obtain the SAR image target recognition result:

[0027] Use the test sample set I test as the input of the trained generative adversarial network model O * and the discriminator network D * makes a class prediction for each test sample I v to obtain the ξ-dimensional class prediction vector D v of I * (I v ), and searches the target class label index table through the subscript corresponding to the maximum value in D * (I v ) to obtain the target class label corresponding to I v .

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] 1. The generator network G in the generative adversarial network model constructed by the present invention includes a plurality of generator basic blocks, and the discriminator network D includes a plurality of discriminator basic blocks. During the training process of the model, the plurality of generator basic blocks and the plurality of discriminator basic blocks can obtain more complete target structure information, improving the target recognition accuracy of the SAR image target recognition method based on generalized principal component analysis and generative adversarial network.

[0030] 2. The present invention uses the optimal projection matrix obtained by performing generalized principal component analysis on the training sample set as part of the weight parameters of the generator network G and the discriminator network D. Since the optimal projection matrix is the optimal representation of the azimuth information of the training samples, the generator network G can generate the azimuth angle target pseudo-images missing in the training sample set, and the discriminator network D can obtain a complete description of the target azimuth information, thereby improving the target recognition accuracy of the SAR image target recognition method based on generalized principal component analysis and generative adversarial network. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is the implementation flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] Referring to Figure 1 , the present invention includes the following steps:

[0034] Step 1) Obtain the training sample set and the test sample set:

[0035] Step 1a) Obtain P SAR images T1 = {T 11 , T 12 , …, T 1p , …, T 1P} at the synthetic aperture radar (SAR) working pitch angle of α, which contain ξ target categories and have target labels, and Q SAR images T2 = {T 21 , T 22 , …, T 2q , …, T 2Q} at the working pitch angle of β, where ξ ≥ 2, T 1p represents the p-th SAR image at the pitch angle of α and containing R α targets, R α ≥ 1, P ≥ 20ξR α , T 2q represents the q-th SAR image at the pitch angle of β and containing R β targets, R β ≥ 1, Q ≥ 10ξR β , Since the working pitch angles of the SAR in the image acquisition stage and the target recognition stage are not always exactly the same, and there is a certain difference between the two, in this example, α = 17°, β = 15°, ξ = 10, R α = 55, R β = 50, P = 50, Q = 50;

[0036] Step 1b) Use a cropping box of size m × n to perform central cropping on the R 1p targets contained in each SAR image T α , and the R 2p targets contained in each SAR image T β respectively, and normalize each cropped image patch. Then, form a training sample set with U normalized image patches I trn = {I1, I2, …, I u , … I U} corresponding to T1, and form a test sample set with V normalized image patches I test = {I1, I2, …, I v , … I V} corresponding to T2, where I u represents the u-th training sample, and I v represents the v-th test sample. In this example, central cropping means taking the target center point as the center of the cropping box and then cropping the target. Normalization uses the maximum-minimum normalization method, and the normalization formula is: I min represents the minimum value of the elements of matrix I, I maxDenote the maximum value of the elements of matrix I, where m = n = 128, U = 2747, and V = 2425;

[0037] Step 2) Perform generalized principal component analysis on the training sample set:

[0038] Since SAR images are a type of exponential family data that follows a non-Gaussian distribution, generalized principal component analysis can project them to obtain a reduced-dimensional representation of the data. The optimal projection matrix extracted is the optimal representation of the azimuth information of the training samples, and at the same time, it can highlight the differences between targets.

[0039] Step 2a) Use a sliding window of size k1×k2 to perform single-step sliding window value extraction on each sample I trn ={I1, I2, …, I u , … I U} in the training sample set I u to obtain the image sub-block matrix X u corresponding to I , and cascade the U image sub-block matrices corresponding to I u to obtain a joint matrix X of size trn , where k1 ≥ 3, k2 ≥ 3,

[0040] denotes rounding up. In this example, k1 = k1 = 9, Step 2b) Initialize an identity matrix Λ of size L×L L , and a projection matrix W of size (k1×k2)×L as a matrix to be determined, and W T W = Λ L , and the generalized principal component projection reconstruction function is and perform S iterations to solve for the matrix W to be determined in F(X, W) to obtain the optimal projection matrix set , where L ≥ 1, S ≥ 2, denotes the s-th optimal projection matrix of size k1×k2×L, μ represents the average value of the elements of the joint matrix X, and W T denotes the transpose operation on W, denotes the operation of taking the two-norm, and the optimal projection matrix W s * is the optimal representation of the azimuth information of the training samples. In this example, L = 8, S = 4;

[0041] The steps to perform S iterations to solve for the matrix W to be determined in F(X, W) are as follows:

[0042] Step 2b1) Initialize the iteration count as s, the maximum iteration count as S, S ≥ 2, and set s = 1;

[0043] Step 2b2) Use the Newton method to solve the undetermined matrix W in the generalized principal component projection reconstruction function F(X, W), and through F(X, W) and the constraint condition W T W = Λ L Construct the conditional equation For Solve the approximate root of, and substitute the obtained approximate root W0 into And The partial derivative of W In, obtain Use Through W0,[[]] And Calculate the projection matrix W of size (k1×k2)×L s , and for the obtained projection matrix W s Transform the matrix dimension in row-major order to obtain the s-th optimal projection matrix W of size k1×k2×L s * , and save W s * To the optimal projection matrix set W * In;

[0044] Step 2b3) Judge whether s≥S holds. If so, obtain the optimal projection matrix set Otherwise, let s = s + 1, and execute step (2b2), where Represents the convolution operation.

[0045] Step 3) Construct the generative adversarial network model O:

[0046] Step 3a) Construct the structure of the generative adversarial network model O:

[0047] Construct the generative adversarial network model O including a sequentially connected generative network G and discriminative network D, where the generative network G uses a two-dimensional transposed convolutional neural network composed of a fully connected layer, S generative basic blocks, and a tangent activation function layer connected in cascade. The generative basic block includes a two-dimensional transposed convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected in cascade; the discriminative network D uses a two-dimensional convolutional neural network composed of S discriminative basic blocks, a fully connected layer, and a Softmax classification output layer connected in cascade. The discriminative basic block includes a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer connected in cascade;

[0048] The generation network G has 16 output channels in its fully connected layer; in the generation basic block, the convolutional kernel of the two-dimensional transposed convolutional layer has a size of 9×9, a convolutional stride of 2×2, a padding of 1×1, and the slope of the LeakyReLU activation function is 0.25;

[0049] The discriminative network D has a convolutional kernel of size 9×9, a convolutional stride of 2×2, a padding of 1×1, and a slope of 0.25 for the ReLU activation function in the two-dimensional convolutional layer of its discriminative basic block; the number of output channels in the fully connected layer is the number of target categories ξ. In this example, S = 4 and ξ = 10;

[0050] Step 3b) Define the loss function L of the generation network G G and the loss function L of the discriminative network D D :

[0051] L G = Ef[log(1 - D(G(z)))]

[0052] L D = Ef[logD(I u )] + Ef[log(1 - D(G(z)))]

[0053] where z represents a random sequence, G(·) represents the transposed sampling function of the generation network G, D(·) represents the class prediction function of the discriminative network D, and Ef[·] represents the expectation function. In this example, z is a random sequence of dimension 100 that follows a normal distribution N(0, 1);

[0054] Step 4) Iteratively train the generative adversarial network model O:

[0055] Step 4a) Initialize the weight parameters of the generation network G and the discriminative network D as {W * , θ G}, {W * , θ D} respectively, the number of iterations is e, the maximum number of iterations is E, and let e = 0. At the same time, fix W * not to participate in the iteration, where E ≥ 200, and θ G represents the parameter set of the bias of the two-dimensional transposed convolutional layer and the weights and biases of the fully connected layer in G, and θ D represents the parameter set of the bias of the two-dimensional convolutional layer and the weights and biases of the fully connected layer in D. In this example, E = 500. To ensure that the generation network G and the discriminative network D can maximize the use of the azimuth information of the training samples contained in the optimal projection matrix, so that the generation network G can generate the azimuth target pseudo-images missing in the training sample set, and the discriminative network D can obtain more target azimuth information, so fix W * not to participate in the iteration;

[0056] Step 4b) Combine the random sequence z with the training sample set I trn As the input of the generative adversarial network model O, the generative network G transposes the random sequence z and obtains the pseudo sample G e (z); Discriminant network D outputs pseudo sample G e (z), and each training sample I u Make category prediction and get G e (z), I u The prediction vector D of the ξ-dimensional category e (G e (z))、D e (I u ),in:

[0057] The fully connected layer in the generative network G upsamples the random sequence z to obtain G1(z); the two-dimensional transposed convolution layer in the first generative basic block performs transposed convolution on the output of the fully connected layer to obtain A1(G1(z))=ConvT(G1(z)), the batch normalization layer in the first generative basic block performs batch normalization on the transposed convolution result to obtain A2(G1(z))=BN(A1(G1(z))), the LeakyReLU activation function layer in the first generative basic block activates the batch normalization result to obtain the output A3(G1(z))=LeakyReLU(A2(G1(z))) of the first generative basic block, and performs S-1 generative basic block operations on A3(G1(z)) to obtain G2(z); the tangent activation function layer activates the output result G2(z) of S generative basic blocks to obtain the pseudo sample G e (z); Among them, the S generation basic blocks enable the generation network G to have the ability to generate pseudo images of targets with missing azimuth angles in the training sample set;

[0058] The pseudo sample G output by the two-dimensional convolutional layer in the first discriminative basic block in the discriminative network D is e (z), and each training sample I u Perform two-dimensional convolution to obtain B1(G e (z))=Conv(G e (z))、B1(I u )=Conv(I u ), the batch normalization layer in the first discriminant basic block performs batch normalization on the two-dimensional convolution results to obtain B2(G e (z))=BN(B1(G e (z)))、B2(I u )=BN(B1(I u )), the ReLU activation function layer in the first discriminant basic block activates the batch normalization results to obtain the output B3(Ge (z)) = ReLU(B2(G e (z))), B3(I u ) = ReLU(B2(I u )) and perform S - 1 discriminative basic block operations on B3(G e (z)) and B3(I u ) respectively to obtain D1(G e (z)) and D1(I u ); The fully - connected layer downsamples the output results D1(G e (z)) and D1(I u ) respectively; The Softmax classification output layer performs probability predictions on the downsampled results of D1(G (z)) and D1(I e (z)) and D1(I u ) respectively to obtain the ξ - dimensional class prediction vectors D e (G u (z)) and D e (G e (z)) and D e (I u ); Among them, the S discriminative basic blocks enable the discriminative network D to have the ability to obtain complete target structure information;

[0059] Step 4c) Adopt the loss function L of the generative network G G , and calculate the loss value of G through the ξ - dimensional class prediction vector D e (z) of G e (G e (z)) Adopt the loss function L of the discriminative network D D , and calculate the loss value of D through the ξ - dimensional class prediction vector D e (z) of G e (G e (z)) and the ξ - dimensional class prediction vector D u of each training sample I e (I u )

[0060] Step 4d) Adopt the gradient - descent method, and update the parameters θ of the generative network G through the partial derivative of the parameter , and update the parameters θ of the discriminative network D through the partial derivative of the parameter G to obtain the weight parameters of G for this iteration as and the weight parameters of D as the partial derivative of the parameter and obtain the weight parameters of D for this iteration as D , and obtain the weight parameters of G for this iteration as and the weight parameters of D as The generative adversarial network model O e , and the parameter update formula is:

[0061]

[0062]

[0063] Wherein, represents the weight parameters of the updated generative network G, represents the weight parameters of G before update, represents the weight parameters of the discriminative network D after update, represents the weight parameters of D before update. In this example, the gradient descent step size λ G = λ D = 0.002;

[0064] Step 4e) Judge whether e≥E holds. If so, obtain the trained generative adversarial network model O * , otherwise, let e = e + 1, O e = O, and execute step (4b);

[0065] Step 5) Obtain the SAR image target recognition result:

[0066] Use the test sample set I test as the input of the trained generative adversarial network model O * , and the discriminative network D * performs class prediction on each test sample I v to obtain the ξ-dimensional class prediction vector D v (I * ), and search the target class label index table through the subscript corresponding to the maximum value in D v (I * ) to obtain the target class label corresponding to I v . v

[0067] The technical effects of the present invention will be further described below in conjunction with simulation experiments:

[0068] 1. Experimental conditions and content:

[0069] The hardware platform for the simulation experiment is: the processor is CPU intel Xeon E5-1630 v4, the main frequency is 3.70GHz, the GPU is NVIDIA GeForce RTX 2080Ti, and the memory is 64GB.

[0070] ​The software platform for the simulation experiment is: Windows 10 Home Edition, 64-bit operating system, Python 3.6, and the PyTorch version is 1.3.0.

[0071] The MSTAR dataset is used for the simulation. This dataset contains target image patches of size 128×128 and corresponding labels for 10 types of targets, namely BMP2 infantry fighting vehicles, BTR70 armored personnel carriers, T72 main battle tanks, 2S1 self-propelled howitzers, BRDM2 armored reconnaissance vehicles, D7 military bulldozers, BTR60 armored personnel carriers, T62 main battle tanks, ZIL131 military trucks, and ZSU23 / 4 self-propelled anti-aircraft guns, at multiple azimuth angles. The target image patches with a radar operating elevation angle of 17° and azimuth angles of 0–90°, 0–180°, 0–270°, and 0–360° are used as the training sets respectively to train the generative adversarial network model constructed by the present invention, and four optimal generative adversarial network models are obtained. The target image patches with a radar operating elevation angle of 15° and full azimuth angles are used as the test sets and are input into the four optimal generative adversarial network models respectively to obtain the predicted target class labels.

[0072] To verify the superiority of the method of the present invention, the overall classification accuracy OA and the average classification accuracy AA are used to quantitatively analyze the algorithm performance and evaluate the recognition accuracy of the present invention.

[0073]

[0074]

[0075] A comparative simulation of the recognition accuracy between the present invention and an existing small-sample SAR image target recognition method is carried out, and the results are shown in Table 1.

[0076] 2. Analysis of simulation results:

[0077] Table 1

[0078]

[0079] Combined with Table 1, it can be seen that compared with the existing technical methods, the average classification accuracy AA of the present invention has increased by 2.46%, 4.53%, 3.92%, and 2.27% respectively under the four training data azimuth settings, and the overall classification accuracy OA has increased by 2.11%, 3.13%, 3.47%, and 1.98% respectively, which proves that the present invention can extract more complete target feature information and improve the SAR image target recognition ability at the same time.

[0080] The above simulation experiments show that: in the present invention, the optimal projection matrix obtained by performing generalized principal component analysis on the training sample set is used as a part of the weight parameters of the generation network G and the discriminant network D, and then an optimal generative adversarial network model is obtained through iterative training, so that the generation network G can generate the azimuth target pseudo-images missing in the training sample set, and the discriminant network D can obtain more complete target azimuth information and structural information, solving the problem that the prior art does not effectively utilize the target azimuth information, and improving the target recognition accuracy rate of the SAR image target recognition method based on generalized principal component analysis and generative adversarial network.

Claims

1. A SAR image target recognition method based on generalized principal component analysis and generative adversarial network GAN, characterized in that, It includes the following steps: (1) Obtain a training sample set and a test sample set: (1a) Obtain P SAR images \(T_1 = \{T\) 11 , T 12 , …, T 1p , …, T 1P \} when the synthetic aperture radar (SAR) working pitch angle is \(\alpha\) and contains \(\xi\) target categories with target labels, and Q SAR images \(T_2 = \{T\) 21 , T 22 , …, T 2q , …, T 2Q \} when the working pitch angle is \(\beta\), where \(\xi\geq2\), \(T\) 1p represents the p-th SAR image with a pitch angle of \(\alpha\) and containing \(R\) α targets, \(R\) α \(\geq1\), \(P\geq20\xi / R\) α , \(T\) 2q represents the q-th SAR image with a pitch angle of \(\beta\) and containing \(R\) β targets, \(R\) β \(\geq1\), \(Q\geq10\xi / R\) β ; (1b) Each SAR image T is cropped with a cropping frame of size m×n 1p for the R α targets it contains, and for each of the R 2p targets contained in each SAR image T β center cropping is performed respectively, and each cropped image patch is normalized. Then, the U normalized image patches I trn corresponding to T1, i.e., I u ={I1, I2, …, I U}, …, I test} form the training sample set, and the V normalized image patches I v corresponding to T2, i.e., I V ={I1, I2, …, I u} form the test sample set. Here, m≥100, n≥100, I v represents the u-th training sample, and I v represents the v-th test sample; (2) Perform generalized principal component analysis on the training sample set: (2a) Take a sliding window of size \(k_1\times k_2\) over the training sample set \(I\) trn =\(\{I_1, I_2, \ldots, I\) u , \ldots I\) U \}\) and take a single-step sliding window value for each sample \(I\) u to obtain the image sub-block matrix \(X\) u corresponding to \(I\), with size . Then concatenate the \(U\) image sub-block matrices corresponding to \(I\) u to obtain the joint matrix \(X\) with size trn , where \(k_1\geq3\), \(k_2\geq3\), and represents rounding up. (2b) Initialize the identity matrix Λ of size L×L L , the projection matrix W of size (k1×k2)×L is a matrix to be determined, and W T W = Λ L , the generalized principal component projection reconstruction function is and perform S iterations to solve the matrix W to be determined in F(X,W) to obtain the optimal projection matrix set where, L≥1, S≥2, represents the s-th optimal projection matrix of size k1×k2×L, μ represents the element average value of the joint matrix X, W T represents the transpose operation of W, represents the operation of taking the two-norm; (3) Construct a generative adversarial network model O: (3a) Construct the structure of the generative adversarial network model O: Construct a generative adversarial network model O including a generative network G and a discriminative network D connected in sequence; wherein the generative network G adopts a two-dimensional transposed convolutional neural network composed of a fully connected layer, S generative basic blocks, and a tangent activation function layer connected in sequence. The generative basic block includes a two-dimensional transposed convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected in sequence; the discriminative network D adopts a two-dimensional convolutional neural network composed of S discriminative basic blocks, a fully connected layer, and a Softmax classification output layer connected in sequence. The discriminative basic block includes a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer connected in sequence; (3b) Define the loss function $L$ of the generation network $G$ G and the loss function $L$ of the discriminator network $D$ D : L G = Ef[log(1 - D(G(z)))] L D = Ef[logD(I u )]+Ef[log(1 - D(G(z)))] wherein, z represents a random sequence, G(·) represents the transposed sampling function of the generative network G, D(·) represents the class prediction function of the discriminative network D, and Ef[·] represents the expectation function; (4) Iteratively train the generative adversarial network model O: (4a) Initialize the weight parameters of the generation network G and the discriminant network D as {W * , θ G} and {W * , θ D} respectively. The number of iterations is e, the maximum number of iterations is E, and let e = 0. At the same time, fix W * not to participate in the iteration. Among them, E ≥ 200, θ G represents the parameter set of the bias of the two-dimensional transposed convolutional layer and the weights and biases of the fully connected layer in G, and θ D represents the parameter set of the bias of the two-dimensional convolutional layer and the weights and biases of the fully connected layer in D; (4b) The random sequence z and the training sample set I trn are used as the inputs of the generative adversarial network model O. The generative network G performs transposed sampling on the random sequence z to obtain the pseudo-sample G e (z); the discriminative network D performs class prediction on the pseudo-sample G e (z) output by G, and each training sample I u to obtain the ξ-dimensional class prediction vectors D e (G u (z)), D e (I e ); e (I u ); (4c) Adopt the loss function \(L\) of the generation network \(G\). G and calculate the loss value of \(G\) through the \(\xi\)-dimensional class prediction vector \(D\) e (G e (z)) e Adopt the loss function \(L\) of the discriminant network \(D\). D and calculate the loss value of \(D\) through the \(\xi\)-dimensional class prediction vector \(D\) e (G e (z)) and the \(\xi\)-dimensional class prediction vector \(D\) e (I u ) of each training sample \(I\) e (I u ) (4d) Using the gradient descent method, and by Parameters Partial derivative of Update the parameters θ of the generative network G G ,pass Parameters Partial derivative of Update the parameters θ of the discriminant network D D , the weight parameter of this iteration G is The weight parameters of and D are Generative Adversarial Network Model O e ; (4e) Determine whether e≥E holds. If so, obtain the trained generative adversarial network model O * , otherwise, set e = e + 1, O e = O, and execute step (4b); (5) Obtain the SAR image target recognition result: Use the test sample set I test as the input of the trained generative adversarial network model O * for the discriminative network D * to perform class prediction on each test sample I v and obtain the ξ-dimensional class prediction vector D v (I * ). Then, find the target class label index table through the subscript corresponding to the maximum value in D v (I * ) and obtain the target class label corresponding to I v . v ​ 2. The SAR image target recognition method based on generalized principal component analysis and generative adversarial network according to claim 1, characterized in that The S - time iterative solution of the undetermined matrix W in F(X,W) described in step (2b) is realized as follows: (2b1) Initialize the iteration number as s, the maximum iteration number as S, S≥2, and let s = 1; (2b2) Use the Newton method to solve the undetermined matrix W in the generalized principal component projection reconstruction function F(X, W), and obtain the projection matrix W of size (k1×k2)×L obtained by the solution s Transform the matrix dimension in row-major order to obtain the s-th optimal projection matrix W of size k1×k2×L s * , and Save it to the optimal projection matrix set W * ; Determine whether s≥S holds. If so, obtain the optimal projection matrix set Otherwise, let s = s + 1, and execute step (2b2), where represents the convolution operation.

3. The SAR image target recognition method based on generalized principal component analysis and generative adversarial network according to claim 1, characterized in that, For the structure of the generative adversarial network model O described in step (3a), the number S of generative basic blocks in the generative network G it contains and the number S of discriminative basic blocks in the discriminative network D are both 4, where: For the generative network G, the output channel number of the fully connected layer it contains is 16; in the generative basic block, the size of the convolution kernel of the two-dimensional transposed convolutional layer is 9×9, the convolution stride is 2×2, the padding is 1×1, and the slope of the LeakyReLU activation function is 0.25; For the discriminative network D, in the discriminative basic block it contains, the size of the convolution kernel of the two-dimensional convolutional layer is 9×9, the convolution stride is 2×2, the padding is 1×1, and the slope of the ReLU activation function is 0.25; the output channel number of the fully connected layer is the number ξ of target classes.

4. The SAR image target recognition method based on generalized principal component analysis and generative adversarial network according to claim 1, characterized in that, The generation network G described in step (4b) performs transposed sampling on the random sequence z, and the discriminant network D performs class prediction on the pseudo-samples G e (z) output by G, as well as each training sample I u The implementation steps are as follows: The fully connected layer in the generation network G upsamples the random sequence z; S generation basic blocks perform two-dimensional transposed convolution on the upsampling result G1(z) of the fully connected layer; the tangent activation function layer activates the result G2(z) of the two-dimensional transposed convolution to obtain the pseudo-sample G e (z); The S discriminative basic blocks in the discriminative network D perform two-dimensional convolutions on the fake samples G e (z) output by G, and each training sample I u respectively; the fully connected layer performs downsampling on the two-dimensional convolution results D1(G e (z)) and D1(I u ) of G e (z) and I u respectively; the Softmax classification output layer performs probability predictions on the downsampling results of D1(G e (z)) and D1(I u ) to obtain the prediction vectors D e (G u (z)) and D e (I e ) of ξ-dimensional categories of G e (z) and I u .

Images