CN106951921B - SAR target recognition method based on Bayesian multi-kernel learning support vector machine - Google Patents

Abstract

The invention discloses a SAR target recognition method based on a Bayesian multi-kernel learning support vector machine, which mainly solves the problem that the existing target recognition methods are inaccurate in SAR image target recognition. The implementation steps are: 1) input the original SAR image and preprocess it, and calculate the kernel matrix of different features; 2) combine the kernel matrix according to the multi-kernel learning method; 3) establish a Bayesian multi-kernel for the support vector machine according to the combined kernel matrix Learning the support vector machine model; 4) using the expectation maximization algorithm to solve the Bayesian multi-kernel learning support vector machine model to obtain the optimal solution; 5) using the optimal solution to perform target recognition on the SAR image test data. The invention effectively combines the inference ability of the Bayesian method and the discrimination ability of the multi-kernel learning method, improves the recognition performance, and can be used for classifying SAR images.

Description

SAR target recognition method based on Bayesian multi-kernel learning support vector machine

technical field

The invention belongs to the technical field of radar target recognition, in particular to a SAR target recognition method, which can be used for SAR image classification.

Background technique

Synthetic Aperture Radar SAR is an active sensor that uses microwaves for perception. Its imaging is not affected by objective factors such as light and climate, and it can monitor targets all day and in all weathers, both in the civilian and military fields. High utilization value. In addition to the target, the SAR image also contains a large number of clutter, and the SAR image also contains a large number of coherent speckles, which makes the detection, identification and identification of the SAR image very difficult; in addition, due to the different configurations of the SAR targets And the complexity of the environment, it is impossible to obtain training samples in all cases. Therefore, how to improve the recognition performance of SAR targets is an important research direction in radar target recognition.

The SAR target recognition methods are mainly divided into the following categories:

One is the method based on template matching;

The second is a model-based approach;

The third is the classification method based on sparse representation;

The fourth is the method based on classifier design, such as K nearest neighbor classifier, neural network classifier, support vector machine and so on.

The support vector machine is a classification algorithm based on statistical learning theory, which cleverly solves the inner product operation problem in high-dimensional space by introducing a kernel function, making the kernel support vector machine in small samples, nonlinear and Specific advantages are shown in high-dimensional pattern recognition problems.

This kind of classifier that combines single data features with support vector machine is called single-kernel learning support vector machine. SVMs show completely different classification performance. Therefore, single-core learning SVMs can only show the characteristics of a certain data feature, but cannot reflect the correlation between various data features, thus affecting the classification performance of the classifier. This reduces the target recognition rate.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a SAR target recognition method based on the Bayesian multi-kernel learning support vector machine in order to improve the target recognition performance in view of the above-mentioned deficiencies of the prior art.

The technical scheme of the present invention is realized as follows:

1. Technical ideas

The invention combines Bayesian inference and multi-core learning method, and introduces multi-core learning method for the selection of different data features, which has good generalization ability and stronger learning ability; Find the solution to the original problem of the support vector machine. The implementation scheme is: first, preprocess the original SAR image to obtain three data features of image domain, frequency domain and sparse coefficient, and calculate the kernel matrix corresponding to the radial kernel function RBF respectively; Finally, use the combined core matrix of the training data to infer the Bayesian model and obtain the optimal solution, and classify the combined core matrix of the test data. The implementation steps include the following:

(A) SAR image preprocessing steps:

A1) Input an original SAR image: I={i _mn |1≤m≤M, 1≤n≤N}, where i _mn represents the amplitude pixel value of the original SAR image, M represents the number of lines of the SAR image, and N Represents the number of columns of the SAR image;

A2) Perform binary segmentation on the original SAR image I, and calculate the centroid of the SAR image

A3) Circular shift the original SAR image I so that the centroid Move to the center of the image to obtain the registration image I ₁ ;

A4) Logarithmic transformation, median filtering and image interception are sequentially performed on the registration image I ₁ to obtain the image domain feature I ₂ of the SAR image, and the image domain feature I ₂ is column vectorized;

A5) do image interception and two-dimensional Fourier transform to the registration image I ₁ , and move the zero frequency to the center of the image, obtain the frequency domain feature I ₃ , and quantize the frequency domain feature I ₃ column;

A6) Repeat the processes A1) to A4) for the SAR image training set and test set respectively to obtain the training data set T _tr and the test data set T _te of the image domain feature;

A7) Repeat the process A1) to A5) for the SAR image training set and test set respectively to obtain the training data set P _tr and the test data set P _te of the frequency domain feature;

A8) Use the KSVD algorithm to learn the image domain feature training set T _tr to obtain a dictionary D and a sparse coefficient feature training data set S _tr corresponding to T _tr , combine the dictionary D and the image domain feature test data set T _te , use the OMP algorithm to calculate Obtain the sparse coefficient feature test data set S _te ;

(B) Multi-core learning steps:

B1) Using the radial kernel function RBF, combined with the image domain feature training data set T _tr and the test data set T _te , calculate the kernel matrix K _ttr (T _tr , T _tr ) of the image domain feature training data set and the image domain feature test Dataset kernel matrix K _tte (T _tr ,T _te );

B2) Using the radial kernel function RBF, combined with the frequency domain feature training data set P _tr and the test data set P _te , calculate the kernel matrix K _ptr (P _tr , P _tr ) of the frequency domain feature training data set and the frequency domain feature test Dataset kernel matrix K _pte (P _tr ,P _te );

B3) Using the radial kernel function RBF, combined with the sparse coefficient feature training data set S _tr and the test data set S _te , calculate the kernel matrix K _str (S _tr , S _tr ) of the sparse coefficient feature training data set and the sparse coefficient feature test Dataset kernel matrix K _ste (S _tr ,S _te );

B4) Combine the training set kernel matrix and test set kernel matrix of the three features calculated in steps B1) to B3), and use the kernel combination method to calculate and obtain the combined kernel matrix K _tr (V', V _tr ) of the SAR image training set and the combined kernel matrix K _te (V', V _te ) of the test set, where V' represents the base vector set, V _tr represents the SAR image training data set, and V _te represents the SAR image testing data set;

(C) Bayesian inference steps:

C1) use the combined kernel matrix K _tr (V', V _tr ) of the SAR image training set to establish a Bayesian multi-kernel learning support vector machine model;

C2) use the expectation maximization algorithm EM to solve the Bayesian multi-kernel learning support vector machine model, and obtain the optimal solution β' of the Bayesian multi-kernel learning support vector machine model;

C3) Using the optimal solution β' of the Bayesian multi-kernel learning support vector machine model obtained in step C2), combined with the combined kernel matrix K _te (V', V _te ) of the SAR image test set, calculate the target category of the SAR image Label y _te .

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

The invention combines the Bayesian support vector machine model with the multi-core learning method, and proposes a SAR target recognition method based on the Bayesian multi-core learning support vector machine, so that the multi-core learning method is superior to the single-core learning method in selecting data features. It can better reflect the correlation between different data features and significantly improve the target recognition performance.

Description of drawings

Fig. 1 is the realization flow chart of the present invention;

FIG. 2 is a schematic diagram of the preprocessing process of the SAR image in the present invention.

Detailed ways

The implementation steps and effects of the present invention will be further described below in conjunction with the accompanying drawings:

Referring to FIG. 1, the implementation steps of the present invention are as follows.

Step 1: Preprocess the SAR image and calculate the kernel matrix.

1a) Input an original SAR image as shown in Figure 2(a): I={i _mn |1≤m≤M, 1≤n≤N}, where _imn represents the amplitude pixel value of the original SAR image, M represents the number of rows of the SAR image, and N represents the number of columns of the SAR image;

1b) using the variable power Ostu segmentation algorithm to perform binary segmentation on the original SAR image I to obtain the segmented SAR image I';

1c) Do the point product calculation of the segmented SAR image I' and the original SAR image I, and the obtained point-multiplied SAR image is shown in Figure 2(b), and calculate the centroid of the point-multiplied SAR image

In the formula, _i'mn represents the pixel value of the SAR image after dot product;

1d) Circular shift the original SAR image I so that the centroid Move to the center of the image to obtain the registered SAR image I ₁ , as shown in Figure 2(c);

1e) performing image interception on the registered SAR image I ₁ to obtain the intercepted SAR image I ₀ , as shown in FIG. 2(d);

1f) Perform logarithmic transformation on the intercepted SAR image I ₀ to obtain a logarithmic SAR image I"', as shown in Figure 2(e);

1g) perform median filter processing on logarithmic SAR image I"' to obtain image domain feature I ₂ , as shown in Figure 2(f), and quantize the image domain feature I ₂ column;

1h) performing image interception and two-dimensional Fourier transform on the registered SAR image I ₁ , and shifting the zero frequency to the center of the image to obtain the frequency domain feature I ₃ , and quantizing the frequency domain feature I ₃ column;

1i) Repeat 1a) to 1g) for the training set and test set of the original SAR image, respectively, to obtain a training data set T _tr and a test data set T _te of image domain features;

1j) Repeat 1a) to 1h) for the training set and test set of the original SAR image, respectively, to obtain the training data set P _tr and the test data set P _te of the frequency domain feature;

1k) Use the KSVD algorithm to learn the image domain feature training set T _tr to obtain a dictionary D and a sparse coefficient feature training data set S _tr corresponding to T _tr , combine the dictionary D and the image domain feature test data set T _te , Use the OMP algorithm to calculate Obtain the sparse coefficient feature test dataset S _te .

In step 2, multi-core learning calculation is performed on the three SAR image features obtained in step 1.

2a) Using the radial kernel function RBF, combined with the image domain feature training data set T _tr and the test data set T _te , calculate the kernel matrix K _ttr (T _tr , T _tr ) of the image domain feature training data set and the image domain feature test The dataset kernel matrix K _tte (T _tr ,T _te ), where the radial kernel function RBF is expressed as follows:

In the formula, q' and q represent two data points in the same space, K(q', q) represents the calculated radial kernel function value, and σ represents the radial kernel function parameter;

2b) Using the radial kernel function RBF, combining the frequency domain feature training data set P _tr and the test data set P _te , calculate the kernel matrix K _ptr (P _tr , P _tr ) of the frequency domain feature training data set and the frequency domain feature test Dataset kernel matrix K _pte (P _tr ,P _te );

2c) Using the radial kernel function RBF, combined with the sparse coefficient feature training data set S _tr and the test data set S _te , calculate the kernel matrix K _str (S _tr , S _tr ) of the sparse coefficient feature training data set and the sparse coefficient feature test Dataset kernel matrix K _ste (S _tr ,S _te );

2d) Combine the training set kernel matrix of the three features calculated in steps 2a) to 2c), and use the kernel combination method to calculate and obtain the combined kernel matrix K _tr (V', V _tr ) of the SAR image training set:

K _tr (V',V _tr )=η _t K _ttr (T _tr ,T _tr )+η _p K _ptr (P _tr ,P _tr )+η _s K _str (S _tr ,S _tr )

In the formula, η _t represents the combination coefficient of the kernel matrix of the image domain feature data set, and its value is 0.5,

η _p represents the combination coefficient of the kernel matrix of the frequency domain feature data set, and its value is 0.5,

η _s represents the combination coefficient of the sparse coefficient feature kernel matrix; the value is 0.5;

2e) Combine the test set kernel matrix of the three features calculated in steps 2a) to 2c), and use the kernel combination method to calculate and obtain the combined kernel matrix K _te (V', V _te ) of the SAR image test set:

K _te (V′,V _te )=η _t K _tte (T _tr ,T _te )+η _p K _pte (P _tr ,P _te )+η _s K _ste (S _tr ,S _te ).

Step 3, build a Bayesian multi-kernel learning support vector machine model.

3a) Given a sample set Among them, x _l represents the training sample, y _l represents the training label, q represents the sample dimension, L represents the number of samples, and T represents the matrix transpose symbol; the unconstrained expression of the maximum edge classifier of the support vector machine is:

In the formula, the first term is the regular term, and the second term is the penalty term; represents the augmented vector, represents the augmented weight, and γ represents the harmonic parameter;

3b) According to Lagrangian duality, the augmented weight solution of SVM maximum edge classifier is obtained In the formula, α _j represents the Lagrangian coefficient;

Will Substitute into the penalty term, we get: The mapping function φ(·) is introduced into the penalty term, and the penalty term expression after the introduction of the mapping function is obtained:

where β _j =α _j y _j , Get the final penalty term expression:

In the formula, β=(β ₁ ,…,β _j ,…,β _L ), represents the augmented vector with augmented vector The kernel function value of , Δ represents the training sample matrix;

3c) Construct the regular term as Add this regular term to the final penalty term to get the final objective function d(β) as:

In the formula, κ represents the harmonic parameter;

3d) Calculate the exponent of the negative of the regular term in the objective function and define it as a pseudo-prior distribution function:

3e) Calculate the index of the final negative penalty term in the objective function, and define it as a pseudo-likelihood distribution function:

In the formula, y=(y ₁ ,...,y _l ,...,y _L ), Δ' represents the training sample matrix;

3f) According to the calculation results in step 3d) and step 3e), the pseudo-posterior distribution function is obtained:

p(β|y,K(Δ,Δ'))∝p(β)p(y|β,K(Δ,Δ')),

3h) Replace K(Δ,Δ') in step 3f) with the combined kernel matrix K _tr (V', V _tr ) of the SAR image training dataset, and establish a Bayesian multi-kernel learning support vector machine model:

p(β|y,K _tr (V',V _tr ))∝p(β)p(y|β,K _tr (V',V _tr ))

In the formula, K _tr (V',V _tr )=(K _tr (V',v ₁ ),...,K _tr (V',v _l ),...,K _tr (V',v _L )),K _tr (V',v _l ) represents the vector formed by the inner product calculation of the basis set and the training sample, p(y|β,K _tr (V',V _tr )) is the pseudo-likelihood after introducing the combined kernel matrix distribution function, where:

Step 4, solve the Bayesian multi-kernel learning support vector machine model.

4a) Use an integral expression containing λ _l to characterize the pseudo-likelihood distribution function after the combined kernel matrix is introduced in step 3h):

Among them, the pseudo-likelihood distribution function has the following relationship:

p(y _l |β,K _tr (V',v _l ))=∫p(y _l ,λ _l |β,K _tr (V',v _l ))dλ _l

Among them, v _l represents the training sample, λ _l represents the hidden variable, and p(y _l ,λ _l |β,K _tr (V',v _l )) is the pseudo-likelihood distribution function after adding the hidden variable;

4b) According to the pseudo-likelihood distribution function after adding the hidden variable in step 4a), a new variable λ is introduced into the Bayesian multi-kernel learning support vector machine model to obtain a new relational expression:

p(β,λ|y,K _tr (V',V _tr ))∝p(β)p(y,λ|β,K _tr (V',V _tr ))

in,

In the formula, λ represents the latent variable vector, λ=(λ ₁ ,…,λ _l ,…,λ _L );

4c) Obtain the posterior distribution function of λ according to the new relational expression in step 4b):

In the formula,

4d) According to the posterior distribution function of λ obtained in 4c), the conditional posterior distribution of λ _l is obtained:

In the formula, represents the generalized inverse Gaussian distribution, according to and The conversion relationship between the The conditional posterior distribution of :

In the formula, represents the inverse Gaussian distribution, and ~ represents the obedience symbol in the distribution function;

4e) According to 4d) The properties of the conditional posterior distribution and the inverse Gaussian distribution are obtained Expected value of:

4f) According to the new relational expression obtained in step 4b) and obtained in 4e) The expectation value of , uses the expectation maximization algorithm EM to solve the Bayesian multi-kernel learning SVM model, and obtains the expression of the iterative solution β ^(m+1) of the Bayesian multi-kernel learning SVM model:

In the formula, m represents the number of iterations of the mth, I represents the identity matrix, express The mth expected iteration value of ;

4g) Set the maximum number of iterations to M', repeat step 4f), stop the iteration when the number of iterations reaches M', and finally obtain the optimal solution β' of the Bayesian multi-kernel learning SVM model:

Step 5: Calculate and obtain the SAR image target recognition category label.

5a) Using the optimal solution β' of the Bayesian multi-kernel learning support vector machine model obtained in step 4g), combined with the SAR image test data set K _te (V', V _te ), use the following formula to obtain the SAR image target identification label y _te :

y _te =sgn(β' ^T K _te (V',V _te ))

In the formula, sgn(·) represents the sign function.

So far, the classification of the SAR target is completed.

The effect of the present invention is further illustrated by the following experiments on the measured data:

1. Experimental scene and parameters:

The data used in the experiment is the public MSTAR dataset of dynamic and static target acquisition and recognition. In this dataset, the image data of BMP2SN9563, BTR70C71, and T72SN132 models at a pitch angle of 17° are selected as training data, and the image data of 7 models at a pitch angle of 15° are selected as test data. BMP2SN9566 and BMP2SNC21 are variants of BMP2SN9563. It is a variant of T72SN132, and the original image size is 128×128.

The data types and number of samples used in this experiment are shown in Table 1:

Table 1 MSTAR experimental data

The experimental parameters are set as follows:

The size of the preprocessed SAR image is 63×63; the radial kernel parameter σ _t = 1 corresponding to the image domain feature, the radial kernel parameter σ p = 0.1 corresponding to the frequency domain feature, and the radial kernel parameter σ _p = 0.1 corresponding to the sparse coefficient feature. The kernel parameter σ _s = 1, the new harmonic parameter κ = 0.01 in the Bayesian multi-kernel learning SVM model;

2. The content and results of this experiment:

The method of the present invention and other existing five methods are used to classify the three types of MSTAR data sets, the first one is linear support vector machine, the second one is single-core learning support vector machine, and the third one is multi-core learning support vector machine machine, the fourth is Bayesian support vector machine, and the fifth is Bayesian single-kernel learning support vector machine;

The experimental steps of carrying out target recognition with the method of the present invention are as follows:

Firstly, the three types of MSTAR data in the experiment are preprocessed, and the radial kernel function is used to obtain three feature kernel matrices, namely image domain feature kernel matrix, frequency domain feature kernel matrix and sparse coefficient feature kernel matrix;

Then, use the combined kernel method to combine the three feature kernel function matrices to obtain the training data set and test data set of the three types of MSTAR data;

Then, the training data sets of the three types of MSTAR data are substituted into the expression of the optimal solution of the Bayesian multi-kernel learning support vector machine and the expression of the expected value of the hidden variable, and the maximum number of iterations is set, and finally the Bayesian multi-kernel is obtained. Learn the optimal solution of support vector machines;

Finally, according to the optimal solution obtained above and the test data set of the three types of MSTAR data, the target identification label is calculated.

The identification results of the three types of MSTAR data by the method of the present invention are compared with the identification results of the other five methods, as shown in Table 2.

Table 2 Results comparison table of the method of the present invention and other methods for MSTAR three types of data

It can be seen from Table 2 that the recognition rate of the Bayesian multi-kernel learning support vector machine model proposed by the present invention for three types of targets in SAR images is 99.12%, which is significantly improved compared with the results of other methods. The performance of target recognition has been significantly improved.

Claims (4)

1. A SAR target recognition method based on Bayesian multi-kernel learning support vector machine, comprising: (A) SAR image preprocessing and kernel matrix calculation steps: A1) Input an original SAR image: I={i _mn |1≤m≤M, 1≤n≤N}, where i _mn represents the amplitude pixel value of the SAR image, M represents the number of lines of the original SAR image, and N represents the number of columns of the original SAR image; A2) Perform binary segmentation on the original SAR image I, and calculate the centroid of the SAR image A3) Circular shift the original SAR image I so that the centroid Move to the center of the image to obtain the registration image I ₁ ; A4) Logarithmic transformation, median filtering and image interception are sequentially performed on the registered SAR image I ₁ to obtain the image domain feature I ₂ of the SAR image, and the image domain feature I ₂ is column vectorized; A5) do image interception and two-dimensional Fourier transform to the registered SAR image I ₁ , and move the zero frequency to the center of the image, obtain the frequency domain feature I ₃ , and quantize the frequency domain feature I ₃ column; A6) Repeat the processes A1) to A4) for the original SAR image training set and test set respectively to obtain the training data set T _tr and the test data set T _te of the image domain feature; A7) Repeat the process A1) to A5) for the original SAR image training set and test set respectively to obtain the training data set P _tr and the test data set P _te of the frequency domain feature; A8) Use the KSVD algorithm to learn the image domain feature training set T _tr to obtain a dictionary D and a sparse coefficient feature training data set S _tr corresponding to T _tr , combine the dictionary D and the image domain feature test data set T _te , use the OMP algorithm to calculate Obtain the sparse coefficient feature test data set S _te ; (B) Multi-core learning steps: B1) Using the radial kernel function RBF, combined with the image domain feature training data set T _tr and the test data set T _te , calculate the kernel matrix K _ttr (T _tr , T _tr ) of the image domain feature training data set and the image domain feature test Dataset kernel matrix K _tte (T _tr ,T _te ); B2) Using the radial kernel function RBF, combined with the frequency domain feature training data set P _tr and the test data set P _te , calculate the kernel matrix K _ptr (P _tr , P _tr ) of the frequency domain feature training data set and the frequency domain feature test Dataset kernel matrix K _pte (P _tr ,P _te ); B3) Using the radial kernel function RBF, combined with the sparse coefficient feature training data set S _tr and the test data set S _te , calculate the kernel matrix K _str (S _tr , S _tr ) of the sparse coefficient feature training data set and the sparse coefficient feature test Dataset kernel matrix K _ste (S _tr ,S _te ); B4) Combine the training set kernel matrix and test set kernel matrix of the three features calculated in steps B1) to B3), and use the kernel combination method to calculate and obtain the combined kernel matrix K _tr (V', V _tr ) of the SAR image training set and the combined kernel matrix K _te (V', V _te ) of the test set, where V' represents the base vector set, V _tr represents the SAR image training data set, and V _te represents the SAR image testing data set; (C) Bayesian inference steps: C1) use the combined kernel matrix K _tr (V', V _tr ) of the SAR image training set to establish a Bayesian multi-kernel learning support vector machine model; C2) use the expectation maximization algorithm EM to solve the Bayesian multi-core learning support vector machine model, and obtain the optimal solution β' of the Bayesian multi-core learning support vector machine; C3) Using the optimal solution β' of the Bayesian multi-kernel learning support vector machine model obtained in step C2), combined with the combined kernel matrix K _te (V', V _te ) of the SAR image test set, calculate the target recognition of SAR images Category label y _te . 2. The method according to claim 1, wherein in step B4), the combined kernel matrix K _tr (V ', V _tr ) of the SAR image training set and the combined kernel matrix K _te (V ) of the test set are calculated using the kernel combination method. ',V _te ), is to use the combined kernel matrix function to add the training set kernel matrix and test set kernel matrix of the three features obtained from B1) to B3) to obtain the combined kernel matrix K _tr of the SAR image training data set (V',V _tr ) and the combined kernel matrix K _te (V',V _te ) of the SAR image test dataset: K _tr (V',V _tr )=η _t K _ttr (T _tr ,T _tr )+η _p K _ptr (P _tr ,P _tr )+η _s K _str (S _tr ,S _tr ) K _te (V' _, V _te )=η _t K _tte (T _tr ,T _te )+η _p K _pte (P _tr ,P _te )+η _s K _ste (S _tr ,S _te ) In the formula, η _t represents the combination coefficient of the kernel matrix of the image domain feature data set, and its value is 0.5, η _p represents the combination coefficient of the kernel matrix of the frequency domain feature data set, and its value is 0.5, η _s represents the combination coefficient of the sparse coefficient feature kernel matrix; the value is 0.5; The combined kernel matrix function is expressed as follows: In the formula, K'(N, N _v ) represents the combined kernel matrix, η _u represents the combined coefficient, N and N _v represent two sets of data points in the same space, K _u (N, N _v ) represents the kernel matrix to be combined, U represents the number of kernel matrices, and R ⁺ represents the set of positive real numbers. 3. The method according to claim 1, wherein in step C1), the combined kernel matrix K _tr (V ', V _tr ) of the SAR image training set is used to set up a Bayesian multi-kernel learning support vector machine model, and is carried out as follows: C11) Given a sample set Among them, x _l represents the training sample, y _l represents the label, q represents the sample dimension, L represents the number of samples, and T represents the matrix transpose; the unconstrained expression of the SVM maximum edge classifier is: In the formula, the first term is the regular term, and the second term is the penalty term; represents the augmented vector, represents the augmented weight, and γ represents the harmonic parameter; C12) Obtain the augmented weight solution of SVM maximum edge classifier according to Lagrangian duality In the formula, α _j represents the Lagrangian coefficient, and y _j represents the corresponding label of the jth sample; Will Substitute into the penalty term, we get: The mapping function φ(·) is introduced into the penalty term, and the penalty term expression after the introduction of the mapping function is obtained: where β _j =α _j y _j , Get the final penalty term: In the formula, β=(β ₁ ,…,β _j ,…,β _L ), L represents the number of samples, represents the augmented vector with augmented vector The kernel function value of , Δ represents the training sample matrix; C13) Construct the regular term as Add this regular term to the final penalty term to get the final objective function d(β) as: In the formula, κ represents the harmonic parameter; C14) Calculate the exponent of the negative of the regular term in the objective function and define it as a pseudo-prior distribution function: C15) Calculate the index of the final negative penalty term in the objective function, and define it as a pseudo-likelihood distribution function: In the formula, y=(y ₁ ,...,y _l ,...,y _L ), Δ' represents the training sample matrix; C16) According to the calculation results in step C14) and step C15), obtain the pseudo-posterior distribution function: p(β|y,K(Δ,Δ'))∝p(β)p(y|β,K(Δ,Δ')), C17) Replace K(Δ,Δ') in step C16) with the combined kernel matrix K _tr (V',V _tr ) of the SAR image training data set, and establish a Bayesian multi-kernel learning support vector machine model as follows: p(β|y,K _tr (V',V _tr ))∝p(β)p(y|β,K _tr (V',V _tr )) In the formula, K _tr (V',V _tr )=(K _tr (V',v ₁ ),...,K _tr (V',v _l ),...,K _tr (V',v _L )),p (y|β,K _tr (V',V _tr )) is the pseudo-likelihood distribution function after introducing the combined kernel matrix, where: 4. The method according to claim 1 or 3, wherein in step C2), use expectation maximization algorithm EM to solve the Bayesian multi-core learning support vector machine model, obtain the optimal solution β' of the Bayesian multi-core learning support vector machine , proceed as follows: C21) Use an integral expression containing λ _l to characterize the pseudo-likelihood distribution function after introducing the combined kernel matrix: Among them, yl represents the sample label, v _l represents the training sample, K _tr (V', v _{l )} represents the vector formed by the inner product calculation of the basis vector set and the training sample, and λ _l represents the latent variable; C22) According to the above-mentioned new pseudo-likelihood distribution function, a new variable λ is introduced into the Bayesian multi-kernel learning support vector machine model, and a new relational expression is obtained: p(β,λ|y,K _tr (V',V _tr ))∝p(β)p(y,λ|β,K _tr (V',V _tr )) in, In the formula, y represents the set of labels, y=(y ₁ ,…,y _l ,…,y _L ), λ represents the latent variable vector, λ=(λ ₁ ,…,λ _l ,…,λ _L ), L represents Number of samples; C23) Obtain the posterior distribution function of λ according to the new relational expression: In the formula, C24) According to the posterior distribution function of λ obtained in C23), the conditional posterior distribution of λ _l is obtained: In the formula, represents the generalized inverse Gaussian distribution, according to and The conversion relationship between the The conditional posterior distribution of : In the formula, Represents the inverse Gaussian distribution, : represents the obedience symbol in the probability distribution; C25) according to C24) in The properties of the conditional posterior distribution and the inverse Gaussian distribution are obtained Expected value of: E(λ _l ^-1 )=|1-yβ ^T K _tr (V',v _l )| ^-1 , C26) According to the posterior distribution function of λ obtained in C23) and the new relational expression obtained in C22), use the expectation maximization algorithm EM to solve the Bayesian multi-kernel learning SVM model, and obtain the Bayesian multi-kernel learning support The expression for the iterative solution β ^(m′+1) of the vector machine model: In the formula, m' represents the number of iterations of the m'th, I represents the identity matrix, express The m'th expected iteration value of ; C27) Set the maximum number of iterations to M', repeat step C26), stop the iteration when the number of iterations reaches M', and finally obtain the optimal solution β' of the Bayesian multi-kernel learning SVM model: