CN106951921B

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

Info

Publication number: CN106951921B
Application number: CN201710149246.9A
Authority: CN
Inventors: 王英华; 王丽业; 刘宏伟; 陈渤; 文伟
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2017-03-14
Filing date: 2017-03-14
Publication date: 2019-07-02
Anticipated expiration: 2037-03-14
Also published as: CN106951921A

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 Bayes multi-core learning support vector machine

Technical Field

The invention belongs to the technical field of radar target identification, and particularly relates to an SAR target identification method which can be used for SAR image classification.

Background

The synthetic aperture radar SAR is an active sensor for sensing by utilizing microwaves, the imaging of the synthetic aperture radar SAR is not influenced by objective factors such as illumination and climate, targets can be monitored all the day long and all the weather, and the synthetic aperture radar SAR has high utilization value in the civil field and the military field. Besides the target, the SAR image also contains a large amount of clutter, and in addition, the SAR image also contains a large amount of coherent speckles, so that the detection, identification and recognition of the SAR image become very difficult; in addition, due to the different configurations of SAR targets and the complexity of the environment in which they are located, it is not possible to obtain training samples in all cases. Therefore, how to improve the identification performance of the SAR target is an important research direction in radar target identification.

The SAR target identification method mainly comprises the following steps:

the method is based on template matching;

second, a model-based approach;

thirdly, a classification method based on sparse representation;

and fourthly, designing a method based on a classifier, such as a K neighbor classifier, a neural network classifier, a support vector machine and the like.

The support vector machine is a classification algorithm based on a statistical learning theory, and skillfully solves the problem of inner product operation in a high-dimensional space by introducing a kernel function, so that the kernel support vector machine has specific advantages in the problems of small samples, nonlinearity and high-dimensional pattern recognition.

The classifier formed by combining the single data characteristic with the support vector machine is called a single-core learning support vector machine, different data characteristics are selected due to different data characteristic characterization data similarity and distinguishing capability, and the single-core learning support vector machine can show completely different classification performances, so that the single-core learning support vector machine can only show the characteristic of a certain data characteristic and cannot show the relevance among the data characteristics, the classification performance of the classifier is influenced, and the target recognition rate is reduced.

Disclosure of Invention

The invention aims to provide an SAR target recognition method based on a Bayesian multi-core learning support vector machine aiming at the defects of the prior art so as to improve the target recognition performance.

The technical scheme of the invention is realized as follows:

technical thought

The invention combines Bayesian inference with a multi-core learning method, introduces the multi-core learning method aiming at the selection problem of different data characteristics, and has good generalization capability and stronger learning capability; meanwhile, Bayesian inference is used for reasoning out the solution of the original problem of the support vector machine. The implementation scheme is as follows: firstly, preprocessing an original SAR image to obtain three data characteristics of an image domain, a frequency domain and a sparse coefficient, and respectively calculating a kernel matrix corresponding to a radial kernel function RBF; thirdly, combining the three kernel matrixes by using a multi-kernel learning algorithm; finally, a Bayesian model is inferred by using the combined kernel matrix of the training data to obtain an optimal solution, and the combined kernel matrix of the test data is classified, wherein the method comprises the following implementation steps:

(A) SAR image preprocessing step:

A1) inputting an original SAR image: i ═ I_mnM is more than or equal to 1 and less than or equal to M, N is more than or equal to 1 and less than or equal to N, wherein i_mnRepresenting the amplitude pixel value of an original SAR image, wherein M represents the row number of the SAR image, and N represents the column number of the SAR image;

A2) performing binary segmentation on the original SAR image I, and calculating to obtain the centroid of the SAR image

A3) Circularly shifting the original SAR image I to make the center of massMoving to the central position of the image to obtain a registration image I₁；

A4) To the registration image I₁Sequentially carrying out logarithmic transformation, median filtering and image interception to obtain image domain characteristics I of the SAR image₂And image domain feature I₂Carrying out column vectorization;

A5) to the registration image I₁Image interception and two-dimensional Fourier transform are carried out, and zero frequency is shifted to the center of the image to obtain frequency domain characteristics I₃And the frequency domain characteristic I₃Carrying out column vectorization;

A6) respectively repeating the processes A1) -A4) on the SAR image training set and the test set to obtain a training data set T of the image domain characteristics_trAnd test data set T_te；

A7) Respectively repeating the processes A1) -A5) on the SAR image training set and the test set to obtain a training data set P of frequency domain characteristics_trAnd a test data set P_te；

A8) Training set T for image domain features using KSVD algorithm_trLearning to obtain dictionaries D and T_trCorresponding sparse coefficient feature training data set S_trTesting the data set T in conjunction with the dictionary D and the image domain features_teUsing OMP algorithm to calculate and obtain sparse coefficient characteristic test data set S_te；

(B) A multi-core learning step:

B1) training data set T using radial kernel function RBF in combination with image domain features_trAnd test data set T_teAnd calculating to obtain a kernel matrix K of the image domain characteristic training data set_ttr(T_tr,T_tr) Sum image domain feature test dataset kernel matrix K_tte(T_tr,T_te)；

B2) Training a data set P using radial kernel function RBF in combination with frequency domain features_trAnd a test data set P_teAnd calculating to obtain a kernel matrix K of the frequency domain characteristic training data set_ptr(P_tr,P_tr) Sum frequency domain feature test dataset kernel matrix K_pte(P_tr,P_te)；

B3) Training data set S by using radial kernel function RBF and combining sparse coefficient features_trAnd a test data set S_teAnd calculating to obtain a kernel matrix K of the sparse coefficient characteristic training data set_str(S_tr,S_tr) And sparse coefficient feature test data set kernel matrix K_ste(S_tr,S_te)；

B4) Combining the training set kernel matrix and the test set kernel matrix of the three characteristics obtained by calculation in the steps B1) -B3), and calculating by using a kernel combination method to obtain a combined kernel matrix K of the SAR image training set_tr(V',V_tr) And combined kernel matrix K of test set_te(V',V_te) Wherein V' represents a set of basis vectors, V_trRepresenting a SAR image training data set, V_teRepresenting a SAR image test dataset;

(C) bayesian inference steps:

C1) combined kernel matrix K using SAR image training set_tr(V',V_tr) Establishing a Bayesian multi-core learning support vector machine model;

C2) solving the Bayes multi-core learning support vector machine model by using an expectation maximization algorithm EM to obtain an optimal solution β' of the Bayes multi-core learning support vector machine model;

C3) using the optimal solution β' of the Bayesian multi-kernel learning support vector machine model obtained in the step C2), and combining the combined kernel matrix K of the SAR image test set_te(V',V_te) And calculating to obtain the target class label y of the SAR image_te。

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

the invention combines a Bayesian support vector machine model with a multi-core learning method, and provides the SAR target recognition method based on the Bayesian multi-core learning support vector machine, so that the multi-core learning method is superior to a single-core learning method in the aspect of selecting data characteristics, the relevance among different data characteristics can be embodied, and the target recognition performance is obviously improved.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention;

fig. 2 is a schematic diagram of a process for preprocessing an SAR image in the present invention.

Detailed Description

The following steps and effects of the present invention will be further explained with reference to the accompanying drawings:

the implementation steps of the invention are as follows with reference to fig. 1.

Step 1, preprocessing the SAR image and calculating a kernel matrix.

1a) Inputting a raw SAR image as shown in fig. 2 (a): i ═ I_mnM is more than or equal to 1 and less than or equal to M, N is more than or equal to 1 and less than or equal to N, wherein i_mnRepresenting the amplitude pixel value of an original SAR image, wherein M represents the row number of the SAR image, and N represents the column number of the SAR image;

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

1c) performing point multiplication calculation on the segmented SAR image I' and the original SAR image I to obtain a point-multiplied SAR image as shown in FIG. 2(b), and calculating the centroid of the point-multiplied SAR image

In formula (II) to'_mnA pixel value representing the point-multiplied SAR image;

1d) circularly shifting the original SAR image I to make the center of massMoving to the central position of the image to obtain a registered SAR image I₁As shown in FIG. 2 (c);

1e) to the registration SAR image I₁Carrying out image interception to obtain an intercepted SAR image I₀As shown in FIG. 2 (d);

1f) for intercepting SAR image I₀Performing logarithmic transformation to obtain a logarithmic SAR image I' "as shown in FIG. 2 (e);

1g) carrying out median filtering processing on the logarithmic SAR image I' ″ to obtain image domain characteristics I₂As shown in FIG. 2(f), image domain feature I₂Carrying out column vectorization;

1h) to the registration SAR image I₁Image interception and two-dimensional Fourier transform are carried out, and zero frequency is shifted to the center of the image to obtain frequency domain characteristics I₃And the frequency domain characteristic I₃Carrying out column vectorization;

1i) respectively repeating 1a) -1 g) on the training set and the testing set of the original SAR image to obtain a training data set T of the image domain characteristics_trAnd test data set T_te；

1j) Respectively repeating 1a) -1 h) on the training set and the test set of the original SAR image to obtain a training data set P of the frequency domain characteristics_trAnd a test data set P_te；

1k) Training set T for image domain features using KSVD algorithm_trLearning to obtain dictionaries D and T_trCorresponding sparse coefficient feature training data set S_trTesting the data set T in conjunction with the dictionary D and the image domain features_teUsing OMP algorithm to calculate and obtain sparse coefficient characteristic test data set S_te。

And 2, performing multi-core learning calculation on the three SAR image characteristics obtained in the step 1.

2a) Training data set T using radial kernel function RBF in combination with image domain features_trAnd test data set T_teAnd calculating to obtain a kernel matrix K of the image domain characteristic training data set_ttr(T_tr,T_tr) Sum image domain feature test dataset kernel matrix K_tte(T_tr,T_te) Wherein 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 a calculated radial kernel function value, and sigma represents a radial kernel function parameter;

2b) training a data set P using radial kernel function RBF in combination with frequency domain features_trAnd a test data set P_teAnd calculating to obtain a kernel matrix K of the frequency domain characteristic training data set_ptr(P_tr,P_tr) Sum frequency domain feature test dataset kernel matrix K_pte(P_tr,P_te)；

2c) Training data set S by using radial kernel function RBF and combining sparse coefficient features_trAnd a test data set S_teAnd calculating to obtain a kernel matrix K of the sparse coefficient characteristic training data set_str(S_tr,S_tr) And sparse coefficient feature test data set kernel matrix K_ste(S_tr,S_te)；

2d) Combining the training set kernel matrixes of the three characteristics obtained in the steps 2a) to 2c), and obtaining a combined kernel matrix K of the SAR image training set by using a kernel combination method_tr(V',V_tr)：

K_tr(V',V_tr)＝η_tK_ttr(T_tr,T_tr)+η_pK_ptr(P_tr,P_tr)+η_sK_str(S_tr,S_tr)

In the formula, η_tThe combination coefficient of the kernel matrix of the image domain characteristic data set is represented, the value is 0.5,

η_pthe combination coefficient of the nuclear matrix of the frequency domain characteristic data set is expressed, the value is 0.5,

η_srepresenting the combined coefficients of a sparse coefficient feature kernel matrix; the value is 0.5;

2e) combining the test set kernel matrixes of the three characteristics obtained by calculation in the steps 2a) to 2c), and calculating by using a kernel combination method to obtain a combined kernel matrix K of the SAR image test set_te(V',V_te)：

K_te(V',V_te)＝η_tK_tte(T_tr,T_te)+η_pK_pte(P_tr,P_te)+η_sK_ste(S_tr,S_te)。

And 3, constructing a Bayesian multi-core learning support vector machine model.

3a) Given sample setWherein x is_lRepresents a training sample, y_lRepresenting a training index, q representing a sample dimension, L representing a sample number, and T representing a matrix transposition symbol; the unconstrained conditional expression of the maximum edge classifier of the support vector machine is as follows:

in the formula, a first term is a regular term, and a second term is a penalty term;the vector of the augmentation is represented by,representing the augmentation weight, and gamma representing a harmonic parameter;

3b) obtaining an augmented weight solution of a Support Vector Machine (SVM) maximum edge classifier according to Lagrange duality In the formula, α_jRepresenting the lagrangian coefficient;

will be provided withSubstituting into the penalty term to obtain:introducing a mapping function phi (-) into the penalty term to obtain a penalty term expression after the mapping function is introduced:

ream type middle β_j＝α_jy_j，And obtaining a final punishment term expression:

in the formula,β＝(β₁,…,β_j,…,β_L),representing augmented vectorsAnd an augmented vectorKernel function value, Δ tableShowing a training sample matrix;

3c) constructing a regularization term ofAdding the regular term and the final penalty term to obtain a final objective function d (β) which is:

wherein κ represents a harmonic parameter;

3d) the exponent of the negative of the regularization term in the objective function is calculated and defined as the pseudo-prior distribution function:

3e) and calculating the index of the negative number of the final penalty term in the target function, and defining the index as a pseudo likelihood distribution function:

wherein y is (y)₁,…,y_l,…,y_L) And Δ' represents a training sample matrix;

3f) obtaining a pseudo posterior distribution function according to the calculation results in the step 3d) and the step 3 e):

p(β|y,K(Δ,Δ'))∝p(β)p(y|β,K(Δ,Δ'))，

3h) combined kernel matrix K for training data set by SAR image_tr(V',V_tr) Replacing K (delta, delta') in the step 3f), and establishing a Bayesian multi-core 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',v₁),…,K_tr(V',v_l),…,K_tr(V',v_L))，K_tr(V',v_l) Represents the vector formed by inner product calculation of the basis set and the training sample, p (y | β, K)_tr(V',V_tr) Is a pseudo-likelihood distribution function after introducing a combined kernel matrix, wherein:

and 4, solving the Bayesian multi-core learning support vector machine model.

4a) By containing lambda_lThe integral expression of (c) represents the pseudo-likelihood distribution function after introducing the combined kernel matrix in step 3 h):

wherein the pseudo-likelihood distribution function has the following relation:

p(y_l|β,K_tr(V',v_l))＝∫p(y_l,λ_l|β,K_tr(V',v_l))dλ_l

wherein v is_lRepresents a training sample, λ_lRepresenting hidden variables, p (y)_l,λ_l|β,K_tr(V',v_l) Is a pseudo-likelihood distribution function after adding hidden variables;

4b) introducing a new variable lambda into the Bayesian multi-core learning support vector machine model according to the pseudo-likelihood distribution function after the hidden variable is added in the step 4a), and obtaining a new relational expression:

p(β,λ|y,K_tr(V',V_tr))∝p(β)p(y,λ|β,K_tr(V',V_tr))

wherein,

in the formula, λ represents an implicit variable vector, and λ ═ λ₁,…,λ_l,…,λ_L)；

4c) Obtaining a posterior distribution function of lambda according to the new relational expression in the step 4 b):

in the formula,

4d) obtaining lambda from the posterior distribution function of lambda obtained in 4c)_lCondition posterior distribution of (1):

in the formula,represents a generalized inverse Gaussian distribution according toAndthe conversion relationship between the two is obtainedCondition a posteriori ofDistribution:

in the formula,representing an inverse gaussian distribution, representing obedient signs in the distribution function;

4e) according to 4d) inProperty of the conditional posterior distribution and inverse Gaussian distribution ofDesired value of (a):

4f) according to the new relational expression obtained in the step 4b) and the relational expression obtained in the step 4e)The expectation value is solved by using an expectation maximization algorithm EM to obtain an iterative solution β of the Bayesian multi-core learning support vector machine model^(m+1)Expression (c):

wherein m represents the mth iteration number, I represents an identity matrix,to representThe mth expected iteration value of (1);

4g) setting the maximum iteration number as M ', repeating the step 4f), stopping iteration when the iteration number reaches M ', and finally obtaining the optimal solution β ' of the Bayesian multi-core learning support vector machine model:

and 5, calculating to obtain the target identification category label of the SAR image.

5a) Utilizing the optimal solution β' of the Bayesian multi-core learning support vector machine model obtained in the step 4g) in combination with the SAR image test data set K_te(V',V_te) Obtaining the SAR image target identification mark y by using the following formula_te：

y_te＝sgn(β'^TK_te(V',V_te))

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

At this point, classification of the SAR target is completed.

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

1. experimental scenarios and parameters:

the data used in the experiments were published dynamic and static target acquisition and identification MSTAR datasets. In the data set, model image data of BMP2SN9563, BTR70C71, and T72SN132 at a pitch angle of 17 ° are selected as training data, model image data of 7 at a pitch angle of 15 ° are selected as test data, and referred to as BMP2SN9566 and BMP2SNC21 as variants of BMP2SN9563, T72SNs7 and T72SN812 as variants of T72SN132, and the original image size is 128 × 128.

The data types and sample numbers used in this experiment are shown in table 1:

TABLE 1 MSTAR Experimental data

The experimental parameters were set as follows:

the size of the preprocessed SAR image is 63 multiplied by 63; radial kernel parameter sigma corresponding to image domain feature_tFrequency domain feature corresponds to a radial kernel parameter σ as 1_p0.1, radial kernel parameter σ corresponding to sparse coefficient feature_s1, a new harmonic parameter k in the Bayesian multi-core learning support vector machine model is 0.01;

2. the content and the result of the experiment are as follows:

the invention and other existing 5 methods classify the MSTAR three kinds of data sets, wherein the 1 st kind is a linear support vector machine, the 2 nd kind is a single-core learning support vector machine, the 3 rd kind is a multi-core learning support vector machine, the 4 th kind is a Bayes support vector machine, and the 5 th kind is a Bayes single-core learning support vector machine;

the experimental steps for target identification by the method of the invention are as follows:

firstly, preprocessing three types of MSTAR data in an experiment, and calculating by using a radial kernel function to obtain three characteristic kernel matrixes, namely an image domain characteristic kernel matrix, a frequency domain characteristic kernel matrix and a sparse coefficient characteristic kernel matrix;

then, combining the three characteristic kernel function matrixes by using a combined kernel method to obtain a training data set and a test data set of the three types of MSTAR data;

then, respectively substituting the training data sets of the MSTAR three types of data into an expression of the optimal solution of the Bayes multi-core learning support vector machine and an expression of the hidden variable expectation value, setting the maximum iteration times, and finally obtaining the optimal solution of the Bayes multi-core learning support vector machine;

and finally, according to the obtained optimal solution and a test data set of the MSTAR three types of data, calculating to obtain a target identification label.

The results of the identification of the MSTAR class three data by the method of the invention were compared with the results of the other 5 methods, as shown in table 2.

TABLE 2 comparison of results of the inventive method and other methods on the MSTAR three types of data

As can be seen from table 2: the recognition rate of the Bayesian multi-core learning support vector machine model on the SAR image three-class target is 99.12%, compared with the results of other methods, the result is obviously improved, and the performance of the SAR image target recognition is obviously improved by the method.

Claims

1. A SAR target recognition method based on a Bayesian multi-core learning support vector machine comprises the following steps:

(A) SAR image preprocessing and kernel matrix calculation steps:

A1) inputting an original SAR image: i ═ I_mnM is more than or equal to 1 and less than or equal to M, N is more than or equal to 1 and less than or equal to N, wherein i_mnRepresenting the amplitude pixel value of the SAR image, wherein M represents the row number of the original SAR image, and N represents the column number of the original SAR image;

A4) To the registration SAR image I₁Sequentially carrying out logarithmic transformation, median filtering and image interception to obtain image domain characteristics I of the SAR image₂And image domain feature I₂Carrying out column vectorization;

A5) to the registration SAR image I₁Image interception and two-dimensional Fourier transform are carried out, and zero frequency is shifted to the center of the image to obtain frequency domain characteristics I₃And the frequency domain characteristic I₃Carrying out column vectorization;

A6) respectively repeating the processes A1) -A4) on the original SAR image training set and the test set to obtain a training data set T of the image domain characteristics_trAnd test data set T_te；

A7) Respectively repeating the processes A1) -A5) on the original SAR image training set and the test set to obtain a training data set P of frequency domain characteristics_trAnd a test data set P_te；

(B) A multi-core learning step:

B2) Training data set using radial kernel function RBF in combination with frequency domain featuresP_trAnd a test data set P_teAnd calculating to obtain a kernel matrix K of the frequency domain characteristic training data set_ptr(P_tr,P_tr) Sum frequency domain feature test dataset kernel matrix K_pte(P_tr,P_te)；

(C) bayesian inference steps:

C2) solving the Bayes multi-core learning support vector machine model by using an expectation maximization algorithm EM to obtain an optimal solution β' of the Bayes multi-core learning support vector machine;

C3) using the optimal solution β' of the Bayesian multi-kernel learning support vector machine model obtained in the step C2), and combining the combined kernel matrix K of the SAR image test set_te(V',V_te) And calculating to obtain the SAR image target identification category label y_te。

2. The method according to claim 1, wherein a combined kernel matrix K for obtaining the SAR image training set is calculated in step B4) by using a kernel combination method_tr(V',V_tr) And combined kernel matrix K of test set_te(V',V_te) B1) -B3) using a combined kernel matrix functionAdding the training set kernel matrix and the test set kernel matrix of the characteristics to obtain a combined kernel matrix K of the SAR image training data set_tr(V',V_tr) Combined kernel matrix K of SAR image test data set_te(V',V_te)：

K_tr(V',V_tr)＝η_tK_ttr(T_tr,T_tr)+η_pK_ptr(P_tr,P_tr)+η_sK_str(S_tr,S_tr)

K_te(V'_,V_te)＝η_tK_tte(T_tr,T_te)+η_pK_pte(P_tr,P_te)+η_sK_ste(S_tr,S_te)

the combined kernel matrix function is represented as follows:

in the formula, K' (N, N)_v) Representing a combined kernel matrix, η_uRepresenting the combined coefficients, N and N_vTwo sets of data points, K, representing the same space_u(N,N_v) Representing the kernel matrices to be combined, U representing the number of kernel matrices, R⁺Representing a positive real number set.

3. Method according to claim 1, wherein in step C1) a combined kernel matrix K of the SAR image training set is used_tr(V',V_tr) Establishing a Bayesian multi-core learning support vector machine model, and carrying out the following steps:

C11) given sample setWherein x is_lRepresents a training sample, y_lDenotes a label, q denotes a sample dimension, L denotes a sample number, and T denotes a matrix transpose; the unconstrained conditional expression of the SVM maximum edge classifier is as follows:

C12) obtaining an augmented weight solution of a Support Vector Machine (SVM) maximum edge classifier according to Lagrange duality In the formula, α_jRepresenting the Lagrangian coefficient, y_jIndicating that the jth sample corresponds to a label;

ream type middle β_j＝α_jy_j，Obtaining a final penalty term:

in the formula,β＝(β₁,…,β_j,…,β_L) And L represents the number of samples,representing augmented vectorsAnd an augmented vectorΔ represents a training sample matrix;

C13) constructing a regularization term ofAdding the regular term and the final penalty term to obtain a final objective function d (β) which is:

wherein κ represents a harmonic parameter;

C14) the exponent of the negative of the regularization term in the objective function is calculated and defined as the pseudo-prior distribution function:

C15) and calculating the index of the negative number of the final penalty term in the target function, and defining the index as a pseudo likelihood distribution function:

C16) obtaining a pseudo posterior distribution function according to the calculation results in the step C14) and the step C15):

p(β|y,K(Δ,Δ'))∝p(β)p(y|β,K(Δ,Δ'))，

C17) combined kernel matrix K for training data set by SAR image_tr(V',V_tr) Replacing K (delta, delta') in the step C16), and establishing a Bayesian multi-core 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',v₁),…,K_tr(V',v_l),…,K_tr(V',v_L)),p(y|β,K_tr(V',V_tr) Is a pseudo-likelihood distribution function after introducing a combined kernel matrix, wherein:

4. the method according to claim 1 or 3, wherein the expectation-maximization algorithm EM is used for solving the Bayesian multi-core learning support vector machine model in the step C2) to obtain an optimal solution β' of the Bayesian multi-core learning support vector machine, and the method comprises the following steps:

C21) by containing lambda_lThe integral expression of (a) represents the pseudo-likelihood after introducing the combined kernel matrixDistribution function:

wherein, y_lDenotes a sample number, v_lRepresents a training sample, K_tr(V',v_l) Representing vectors, lambda, formed by performing inner product calculation on the basis vector set and the training samples_lRepresenting hidden variables;

C22) according to the new pseudo-likelihood distribution function, introducing a new variable lambda into a Bayesian multi-core 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))

wherein,

wherein y represents a set of reference numerals, and y ═ y₁,…,y_l,…,y_L) λ represents an implicit variable vector, λ ═ λ₁,…,λ_l,…,λ_L) L represents the number of samples;

C23) obtaining a posterior distribution function of lambda according to a new relational expression:

in the formula,

C24) obtaining lambda according to the posterior distribution function of lambda obtained in C23)_lCondition posterior distribution of (1):

in the formula,represents a generalized inverse Gaussian distribution according toAndthe conversion relationship between the two is obtainedCondition posterior distribution of (1):

in the formula,representing an inverse gaussian distribution, representing obedient signs in the probability distribution;

C25) according to C24)Property of the conditional posterior distribution and inverse Gaussian distribution ofDesired value of (a):

E(λ_l ^-1)＝|1-yβ^TK_tr(V',v_l)|^-1，

C26) according to the posterior distribution function of lambda obtained in C23) and the new relational expression obtained in C22), the Bayesian multi-core learning support vector machine model is solved by using an expectation-maximization algorithm EM, and the iterative solution β of the Bayesian multi-core learning support vector machine model is obtained^(m′+1)Expression (c):

wherein m 'represents the m' th iteration number, I represents an identity matrix,to representThe m' th expected iteration value of (a);

C27) setting the maximum iteration number as M ', and repeating the step C26), stopping iteration when the iteration number reaches M ', and finally obtaining the optimal solution β ' of the Bayesian multi-core learning support vector machine model: