CN104217436B - SAR image segmentation method based on multiple features combining sparse graph - Google Patents

Abstract

The invention discloses a SAR image segmentation method based on a multi-feature joint sparse graph, which mainly solves the problems that the existing method has adverse effects on the segmentation results and the computational complexity is affected by the increase of the image scale. The implementation steps are: 1) Input the image to be segmented; 2) Over-segment the image to be segmented to obtain superpixels; 3) Extract four feature sets of superpixels; 4) Jointly sparsely represent the four feature sets to obtain four Sparse representation coefficients; 5) Fusion of four sparse representation coefficients into one global sparse representation coefficient; 6) Computing local spatial neighborhood correlation of superpixels; 7) Combining global sparse representation coefficients with local spatial neighborhood correlation to generate Joint sparse graph; 8) Use the spectral clustering algorithm to segment the vertices of the joint sparse graph to obtain the final segmentation result. The invention can effectively segment the SAR image and is not affected by noise, and can be used for automatic target recognition of the SAR image.

Description

SAR image segmentation method based on multi-feature joint sparse graph

technical field

The invention belongs to the technical field of image processing, in particular to a SAR image segmentation used for SAR image target recognition.

Background technique

Synthetic Aperture Radar (SAR) is a high-resolution imaging radar that can penetrate clouds and vegetation and is almost unaffected by climate conditions. It can work all day and all day, so it is widely used in ground monitoring and identification. Due to the characteristics of the SAR system, SAR images are not as intuitive as optical remote sensing images, so the later understanding and interpretation of SAR images is also extremely critical. SAR image segmentation, as one of the key links in SAR image interpretation, has attracted extensive attention in recent years. Many methods have been used for SAR image segmentation, such as: threshold, clustering, support vector product, Markov random field and so on.

Segmentation methods based on support vector products and Markov random fields require a small number of samples, which are difficult to obtain accurately for SAR images. For the threshold segmentation method, it is usually difficult to choose an appropriate threshold. At present, many mature clustering algorithms have been used for SAR image segmentation, but most of the clustering methods are based on pixels; the complexity of such methods will increase sharply with the increase of image size, thus affecting the clustering. Method promotion and application.

In addition, since the SAR image is obtained through the interference of radar echoes, the SAR image itself inevitably contains coherent speckle noise. As we all know, in SAR image interpretation, the suppression of coherent speckle noise in SAR image is extremely critical. In order to avoid the influence of noise on the segmentation results, most of the existing SAR image segmentation algorithms use filtering algorithm to preprocess the SAR image, and then apply other techniques to segment it. However, while the filtering algorithm reduces the noise, the edges and textures in the image will become blurred, which causes the irreversible loss of some details in the image, which may lead to inaccurate final segmentation results.

Contents of the invention

The purpose of the present invention is to address the deficiencies of the above-mentioned prior art, and propose a SAR image segmentation method based on a joint sparse graph, so as to avoid the influence of coherent speckle noise and parameter selection, and obtain better SAR image segmentation results.

To achieve the above object, the technical solution of the present invention comprises the following steps:

(1) Input the SAR image to be segmented, and determine the number of categories k ≥ 2 that need to be divided into the image;

(2) Use the region growing algorithm to over-segment the input SAR image to be segmented, and generate multiple superpixel sets Y representing homogeneous regions;

(3) For each superpixel in the superpixel set Y, extract its grayscale histogram feature, local binary mode histogram feature, Gabor filter bank feature, and grayscale co-occurrence matrix statistical feature to form a grayscale histogram feature set X ₁ , local binary pattern histogram feature set X ₂ , Gabor filter bank feature set X ₃ , gray level co-occurrence matrix probability statistics feature set X ₄ ;

(4) Jointly sparsely represent the above four feature sets X ₁ , X ₂ , X ₃ , and X ₄ of all superpixels, and obtain the sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , and Z ₄ of the four feature sets respectively , the size of each sparse representation coefficient is n×n, where n is the number of superpixels;

(5) Among the four sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , Z ₄ of size n×n, take any element [Z1] _ij , [Z2] _ij , [Z3] _ij , [ Z4] _ij , among them, 1≤i≤n, 1≤j≤n, these four elements are fused to obtain the elements at the same position in the global sparse representation coefficient Each element of the four sparse representation coefficients is fused according to the above formula to obtain the global sparse representation coefficient

(6) Construct the local spatial neighborhood correlation C of superpixels according to the adjacent relationship of superpixels. The local spatial neighborhood correlation C is a square matrix with the number of superpixels as the number of rows and columns. For any Two superpixels Y _i and Y _j , when Y _i and Y _j are adjacent, the value of the j-th element C _ij in the i-th row of C is 1, otherwise the value of C _ij is 0;

(7) Combine the global sparse representation coefficient S with the local spatial neighborhood correlation C to generate the adjacency matrix G=(1/2)(P+P ^T ) of the joint sparse graph, where P=S·exp(C/ 2σ ² ), T means transpose, σ is a regular parameter with a value of 1.4;

(8) Use the canonical cut Ncuts algorithm in spectral clustering to divide the vertices of the joint sparse graph into k classes, and obtain the final segmentation result.

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

1) In the present invention, there is no need to use two methods to reduce noise and mine the global structure of the image, but only use a joint sparse map based on multiple features of superpixels to achieve the same purpose, so without affecting the computational complexity Under the premise, it not only increases the size of the image, but also avoids the influence of the filtering algorithm on the segmentation results.

2) The present invention enhances the classification ability of the joint sparse graph by utilizing the local spatial neighborhood information of the image, thereby obtaining better image segmentation results.

Description of drawings

Fig. 1 is the overall realization flowchart of the present invention;

Fig. 2 is the comparison figure of the segmentation result of the present invention and existing method to two kinds of SAR images;

Fig. 3 is a comparison diagram of segmentation results of three types of SAR images between the present invention and the existing method.

detailed description

With reference to Fig. 1, the concrete realization steps of the present invention are as follows:

Step 1, input the SAR image to be segmented, judge the main target and background to be recognized according to the image content, and determine the number of segmentation classes k, the values of k in this example are 2 and 3.

Step 2, use the region growing algorithm to generate superpixels.

(2a) For the SAR image to be segmented where all pixels have not been merged into the region, set the first pixel in the upper left corner of the image to be (x0, y0);

(2b) Compare the pixel value of any pixel (x, y) in the 8 neighborhoods centered on the pixel (x0, y0) with the threshold threshold Th=45, if the pixel value of the pixel (x, y) is less than the threshold threshold Th , the pixel (x, y) and (x0, y0) are merged into one area, and the pixel (x, y) is pushed into the stack at the same time;

(2c) Take a pixel from the stack and return it to step (2b) as (x0, y0);

(2d) When the stack is empty, scan the image sequentially, find the first pixel in the SAR image to be segmented that has not been merged into the area, set the pixel as (x0, y0), and return to step (2b);

(2e) Repeat steps (2b) to (2d) until each pixel in the SAR image to be segmented is merged into a certain region to obtain multiple regions;

(2f) Treat each region as a superpixel, and use all regions to form a superpixel set Y.

Step 3, extract four kinds of features of all superpixels to form four feature sets respectively.

SAR images have rich amplitude and texture information. In order to obtain more accurate labels for each superpixel, it is necessary to perform texture analysis on the superpixel set Y before dividing it, and extract four kinds of features of each superpixel. The specific steps are as follows :

(3a) Count the histogram of each superpixel, arrange the histogram features of any superpixel Y _j in a row, and obtain the histogram feature column vector x _1,j of the superpixel Yj, according to the superpixels in the superpixel set Y In order, arrange the histogram feature column vectors of all superpixels in order to form a histogram feature set X ₁ =(x _1,1 ,x _1,2 ,…,x _1,j ,…,x _1,n ), where, 1≤j≤n, n is the number of superpixels;

(3b) For each pixel of the SAR image to be segmented, calculate the local binary mode in its eight neighborhoods, count the local binary mode histogram of each superpixel, and calculate the local binary mode histogram of any superpixel Y _j The graph features are arranged in a column to obtain the local binary mode histogram feature column vector x _2,j of the superpixel Y _j , and according to the order of the superpixels in the superpixel set Y, the local binary mode histogram feature column vectors of all superpixels Arranging in sequence to form a local binary pattern histogram feature set X ₂ =(x _2,1 ,x _2,2 ,...,x _2,j ,...,x _2,n ), where 1≤j≤n;

(3c) Calculate a total of 40 Gabor filters in 5 scales and 8 directions to form a Gabor filter bank, use the Gabor filter bank to filter the SAR image to be segmented, and calculate the filtering results of each filter for each superpixel Arrange the average value of 40 Gabor filter results of any superpixel Y _j in a row to obtain the Gabor filter group feature column vector x _3,j of superpixel Y _j , according to the order of superpixels in superpixel set Y , arrange the Gabor filter bank feature column vectors of all superpixels in order to form a Gabor filter bank feature set X ₃ =(x _3,1 ,x _3,2 ,…,x _3,j ,…,x _3,n ) , where 1≤j≤n;

(3d) Calculate the gray-level co-occurrence matrix of each superpixel, and count 14 kinds of probability and statistical characteristics of the gray-level co-occurrence matrix: angular second-order distance, contrast, correlation, sum of squares, inverse moment, sum of average, sum of variance, entropy, The sum of entropy, variation, variance of entropy, X-direction information correlation, Y-direction information correlation, and maximum correlation coefficient, arrange the 14 kinds of probability and statistical features of any superpixel Y _j in a row, and obtain the gray value of superpixel Y _j Degree co-occurrence matrix probability and statistics feature column vector x _4,j , according to the order of the superpixels in the superpixel set Y, arrange the gray-level co-occurrence matrix probability and statistics feature column vectors of all superpixels in sequence to form the gray-level co-occurrence matrix probability and statistics feature set X ₄ =(x _4,1 ,x _4,2 ,...,x _4,j ,...,x _4,n ), where 1≤j≤n.

Step 4, perform joint sparse representation on the four feature sets of all superpixels.

In order to effectively combine the four features, the present invention utilizes the cross information among the four features, so that the different features of the same superpixel have consistent sparsity. For the above four feature sets X ₁ , X ₂ of all superpixels , X ₃ , and X ₄ are jointly sparsely represented to obtain four sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , and Z ₄ , which are obtained by solving the following formula:

stdiag(Z _f )=0, f=1,...,m,

Among them, X _f =[x _f,1 ,x _f,2 ,…,x _f,j ,…,x _f,n ] represents the fth feature set, and x _f,j is the jth superpixel set Y The characteristics of superpixels, n is the number of superpixels; α>0 and β>0 are two parameters that weigh the effects of each part, and their values are α=0.5, β=0.3, m=4; ||·|| _F is the F-norm of the matrix, is a fidelity item; ||Z _f || _col,1 is the 1 norm of each column of Z _f ; when f=1, the sparse representation coefficient of the feature set X ₁ is Z ₁ , and when f=2, the feature set X The sparse representation coefficient of ₂ is Z ₂ , when f=3, the sparse representation coefficient of feature set X ₃ is Z ₃ , when f=4, the sparse representation coefficient of feature set X ₄ is Z ₄ , Z is four sparse representation coefficients The joint sparse representation coefficient of Z ₁ , Z ₂ , Z ₃ , and Z ₄ has the following specific form:

where ||Z|| _1,2 is the 1,2 norm of the joint sparse representation coefficient Z, Where Z _ij is the jth element of the i-th row in Z, 1≤i≤n, 1≤j≤n.

Step 5. Calculate the four sparse representation coefficients using the Alternating Direction Multiplier Algorithm.

In order to calculate the four sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , Z ₄ , use the alternating direction multiplier algorithm ADMM to solve the formula <1>, and convert the formula <1> into the following equivalent form:

stZ _f =H _f , diag(H _f )=0, Z _f =J _f , f=1,...,m,

Transform the above equivalent form into a Lagrangian function, so that the solution formula <1> becomes the minimization of the following Lagrangian function:

Among them, H ₁ ,…,H _m are the first auxiliary variables, J ₁ ,…,J _m are the second auxiliary variables, U ₁ ,…,U _m are the first Lagrangian multipliers, V ₁ ,…, V _m is the second Lagrangian multiplier, and μ>0 is a penalty parameter. The specific solution steps are as follows:

(5a) Initialize various parameters: set convergence threshold η=10 ^-5 , scaling factor θ=1.1, penalty parameter μ=10 ^-6 , Z _f =H _f =J _f =U _f =V _f =0;

(5b) Use the soft threshold method to update each column of the first auxiliary variable H ₁ ,…,H _m , set h _f,e =(h _f,1e ,h _f,2e ,…,h _f,ne ) ^T is the e-th column of H _f (f=1,...,m), in order to satisfy the constraint condition diag(H _f )=0, remove the e-th element in h _{f, e} to get the auxiliary column vector Then the auxiliary column vector according to the following formula Updates are performed on the updated auxiliary column vector Insert 0 at the e-th position in , and get the e-th column h _f,e in the updated first auxiliary variable H _f =(h _f,1e ,h _f,2e ,…,h _{f,(e-1) e} ,0,h _f,(e+1)e ,…,h _f,ne ) ^T :

in, is the sparse representation coefficient column vector obtained by removing the e-th element of the e-th column of the sparse representation coefficient Zf, Is the Lagrangian multiplier column vector obtained by removing the e-th element of the e-th column of the first Lagrange multiplier U _f ;

(5c) Update the second auxiliary variables J ₁ ,...,J _m according to the following formula:

Wherein, f=1,...,m, I is the identity matrix that size is n * n;

(5d) Update all rows of the joint sparse representation coefficient Z, let z _i be the ith row of Z, update z _i according to the following formula:

Among them, ||q _i || ₂ is the 2-norm of q _i , q _i is the ith row of the intermediate variable Q, and Q is an n ² ×m matrix, and its specific form is as follows:

Wherein, Y _f =(H _f +J _f -(U _f +V _f )/μ)/2, f=1,...,m;

(5e) Update the first Lagrangian multipliers U ₁ ,…,U _m and the second Lagrangian multipliers V ₁ ,…,V _m , and the penalty parameter μ according to the following formula:

U _f =U _f +μ(Z _f -H _f ), f=1,...,m,

V _f =V _f +μ(Z _f -J _f ), f=1,...,m,

μ=min(θμ,10 ¹⁰ );

(5f) Convert each n ² ×1 column vector of the joint sparse representation coefficient Z into an n×n matrix, then the first column of Z is converted into matrix Z ₁ , and the second column of Z is converted into matrix Z ₂ , the third column of Z is converted into matrix Z ₃ , and the fourth column of Z is converted into matrix Z ₄ , thereby obtaining four sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , Z ₄ ;

(5g) Compare the infinite norm of Z _f −H _f ||Z _f −H _f || _∞ and the infinite norm of Z _f −J _f ||Z _f −J _f || _∞ respectively with the convergence threshold η , if both ||Z _f -H _f || _∞ and ||Z _f -J _f || _∞ are smaller than the convergence threshold η, then obtain four sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , Z ₄ at the same time, otherwise Return to step (5b).

Step 6, fusing the four sparse representation coefficients to obtain a global sparse representation coefficient.

In order to obtain a unified global sparse representation coefficient, the sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , and Z ₄ of the four features need to be fused, and the four n×n sparse representation coefficients Z ₁ , Z ₂ , In Z ₃ , Z ₄ , take any elements [Z1] _ij , [Z2] _ij , [Z3] _ij , [Z4] _ij in turn at the same position, among them, 1≤i≤n, 1≤j≤n, the The four elements are fused according to the following formula to obtain the elements at the same position in the global sparse representation coefficient:

Each element of the four sparse representation coefficients is fused according to the above formula to obtain the global sparse representation coefficient

Step 7, calculate the local spatial neighborhood correlation of superpixels.

The global sparse representation coefficient S only reflects the global similarity of superpixels. In order to comprehensively explore the relationship between superpixels, the present invention also mines the local spatial neighborhood information of superpixels, that is, constructs superpixels in a simple way. Local spatial neighborhood correlation C, the local spatial neighborhood correlation C is a square matrix with the number of superpixels as the number of rows and columns, for any two superpixels Y _i and Y _j , when Y _i and Y When _j is adjacent, the value of the jth element C _ij in the i row in C is 1, otherwise the value of C _ij is 0.

Step 8, generate a joint sparse graph.

The present invention combines the global joint sparse representation coefficient S with the local spatial neighborhood correlation C according to the following formula:

P=S·exp(C/2σ ² ),

Among them, σ is a regular parameter, and P is the adjacency matrix of a directed graph. In order to maintain the stability of information spreading between the vertices of the graph, it is usually transformed into a joint sparse graph. The adjacency matrix of the joint sparse graph is A joint sparse graph embodying global structure and local structure is thus obtained.

Step 9, use the Ncuts algorithm in spectral clustering to segment the joint sparse graph.

Based on the spectral graph theory in graph theory, spectral clustering transforms the segmentation problem of superpixel set Y into the optimal partition problem of the joint sparse graph corresponding to superpixel set Y. In the present invention, the elements in the superpixel set Y are divided into multiple categories using the canonical cut Ncuts algorithm. Taking two types of problems as an example, the superpixel set Y is divided into two disjoint subsets A and B:

in, G _ab is the b-th element of row a in the joint sparse graph adjacency matrix G, G _ay is the y-th element in the a-th row of the joint sparse graph adjacency matrix G, G _by is the b-th element in the joint sparse graph adjacency matrix G The yth element of the row.

Effect of the present invention can be verified by following simulation experiments:

1. Experimental condition setting

In the present invention, the number of categories k of SAR images is divided into 2 and 3, which are verified on two SAR images respectively.

When k=2, the second type of SAR image to be segmented is taken by RADARSAT satellite, which reflects the surface changes in the Ottawa area of Canada affected by the rainy season. The shooting time is August 1997, and the size is 313×352. It includes land and water The two parts, due to the different heights of the land part, correspond to very similar gray values on the SAR image, and it is difficult to divide the two regions accurately by general segmentation methods.

When k=3, the three types of SAR images to be segmented are Ku-band SAR images near the Rio Grande River in the United States, with a size of 200×180, which includes water, cultivated land and vegetation, due to the contrast difference between cultivated land and vegetation Smaller, vegetation is difficult to distinguish.

The over-segmentation threshold Th=45 used in the present invention, the parameters α=0.5, β=0.3, m=4 in the formula <1>, and the joint sparse graph regularization parameter σ=1.4.

2. Simulation content and results

A) Using the existing low-rank map, L1 map and the method of the present invention to carry out simulation experiments on two types of SAR images, the segmentation results are shown in Figure 2. Among them, Figure 2(a) is the second-class SAR image, Figure 2(b) is the segmentation result of the second-class SAR image based on the low-rank image, Figure 2(c) is the segmentation result of the second-class SAR image based on the L1 image, and Figure 2( d) is the segmentation result of the second-class SAR image of the present invention.

It can be seen from Fig. 2(b) that although the segmentation result of the second-class SAR image based on the low-rank map generally shows the outline of the water area, there are many noise points in the land part.

It can be seen from Figure 2(c) that the classification results of Class II SAR images based on the L1 map have better regional consistency than those based on the low rank map, but the land parts around the water area have Lots of misclassifications.

It can be seen from Fig. 2(d) that due to the combination of the global structure and the local structure in the present invention, the segmentation result of the second-class SAR image clearly presents two parts of water and land, and each part has a good regional consistency, which is It also meets the needs of ground monitoring.

B) Using the existing low-rank map, L1 map and the present invention to conduct simulation experiments on three types of SAR images, the segmentation results are shown in FIG. 3 . Among them, Figure 3(a) is the three types of SAR images, Figure 3(b) is the segmentation results of three types of SAR images based on low-rank images, Figure 3(c) is the segmentation results of three types of SAR images based on L1 images, Figure 3( d) is the segmentation result of three types of SAR images in the present invention.

It can be seen from Fig. 3(b) that although the segmentation results of the three types of SAR images based on the low-rank map segmented a small amount of vegetation, it did not distinguish large areas of water and cultivated land.

It can be seen from Fig. 3(c) that the segmentation results of the three types of SAR images based on the L1 map can basically separate the water area, vegetation and cultivated land compared with the results of the three types of SAR image segmentation based on the low-rank image. There is misclassification in the vegetation part.

It can be seen from FIG. 3( d ) that the water area, vegetation and cultivated land are more accurately divided in the three types of SAR image segmentation results of the present invention.

Claims (5)

1. The SAR image segmentation method based on multi-feature joint sparse graph, comprises the following steps: (1) Input the SAR image to be segmented, and determine the number of categories k ≥ 2 that need to be divided into the image; (2) Use the region growing algorithm to over-segment the input SAR image to be segmented, and generate multiple superpixel sets Y representing homogeneous regions; (3) For each superpixel in the superpixel set Y, extract its grayscale histogram feature, local binary mode histogram feature, Gabor filter bank feature, and grayscale co-occurrence matrix statistical feature to form a grayscale histogram feature set X ₁ , local binary pattern histogram feature set X ₂ , Gabor filter bank feature set X ₃ , gray level co-occurrence matrix statistical feature set X ₄ ; (4) Jointly sparsely represent the above four feature sets X ₁ , X ₂ , X ₃ , and X ₄ of all superpixels, and obtain the sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , and Z ₄ of the four feature sets respectively , the size of each sparse representation coefficient is n×n, where n is the number of superpixels; (5) In the four sparse representation coefficients Z ₁ , Z ₂ , Z ₃ , Z ₄ of size n×n, respectively take any element [Z ₁ ] _ij , [Z ₂ ] _ij , [Z ₃ ] at the same position _ij ,[Z ₄ ] _ij , among them, 1≤i≤n, 1≤j≤n, these four elements are fused to obtain the elements at the same position in the global sparse representation coefficient Each element of the four sparse representation coefficients is fused according to the above formula to obtain the global sparse representation coefficient (6) Construct the local spatial neighborhood correlation C of superpixels according to the adjacent relationship of superpixels. The local spatial neighborhood correlation C is a square matrix with the number of superpixels as the number of rows and columns. For any Two superpixels Y _i and Y _j , when Y _i and Y _j are adjacent, the value of the j-th element C _ij in the i-th row of C is 1, otherwise the value of C _ij is 0; (7) Combine the global sparse representation coefficient S with the local spatial neighborhood correlation C to generate the adjacency matrix G=(1/2)(P+P ^T ) of the joint sparse graph, where P=S·exp(C/ 2σ ² ), T means transpose, σ is a regular parameter with a value of 1.4; (8) Use the canonical cut algorithm Ncuts in spectral clustering to divide the vertices of the joint sparse graph into k classes to obtain the final segmentation result. 2. the SAR image segmentation method based on the multi-feature joint sparse graph according to claim 1, wherein in the step (3), the gray histogram feature set X ₁ is formed as follows: (3a) Count the histogram of each superpixel, arrange the histogram features of any superpixel Y _j into a column, and obtain the histogram feature column vector x1 _{,j of the superpixel Y j ;} (3b) According to the order of the superpixels in the superpixel set Y, arrange the histogram feature column vectors of all superpixels in order to form a histogram feature set X ₁ =(x _1,1 ,x _1,2 ,…,x _1,j ,…,x _1,n ), where, 1≤j≤n, n is the number of superpixels. 3. the SAR image segmentation method based on multi-feature joint sparse graph according to claim 1, wherein in the said step (3), the local binary pattern histogram feature set X ₂ is formed as follows: (3c) For each pixel of the SAR image to be segmented, calculate the local binary pattern in its eight neighborhoods, and count the local binary pattern histogram of each superpixel; (3d) arranging the local binary pattern histogram features of any one superpixel Y _j into a column to obtain the local binary pattern histogram feature column vector x _{2,j of the superpixel Y j ;} (3e) According to the order of the superpixels in the superpixel set Y, arrange the local binary pattern histogram feature column vectors of all superpixels in sequence to form the local binary pattern histogram feature set X ₂ =(x _2,1 ,x _{2, 2} ,…,x _2,j ,…,x _2,n ), where, 1≤j≤n, n is the number of superpixels. 4. the SAR image segmentation method based on multi-feature joint sparse graph according to claim 1, wherein in said step (3), Gabor filter bank feature set X ₃ , constitutes as follows: (3f) Calculate a total of 40 Gabor filters in 5 scales and 8 directions to form a Gabor filter bank, use the Gabor filter bank to filter the SAR image to be segmented, and calculate the filtering results of each filter for each superpixel average value; (3g) arrange the 40 Gabor filtering result mean values of any superpixel Y _j into a row, and obtain the Gabor filter group feature column vector x _{3,j of the superpixel Y j ;} (3h) According to the order of the superpixels in the superpixel set Y, arrange the Gabor filter bank feature column vectors of all superpixels in turn to form the Gabor filter bank feature set X ₃ =(x _3,1 ,x _3,2 ,…, x _3,j ,…,x _3,n ), where, 1≤j≤n, n is the number of superpixels. 5. the SAR image segmentation method based on the multi-feature joint sparse graph according to claim 1, wherein in the step (3), the gray level co-occurrence matrix statistical feature set X ₄ is formed as follows: (3i) Calculate the gray-level co-occurrence matrix of each superpixel, and count 14 kinds of probability and statistical characteristics of the gray-level co-occurrence matrix: angular second-order distance, contrast, correlation, sum of squares, inverse moment, sum of average, sum of variance, entropy, Entropy sum, variation, variance of entropy, information correlation in X direction, information correlation in Y direction, maximum correlation coefficient; (3j) Arranging 14 kinds of probability and statistical features of any superpixel Y _j in a row to obtain the gray level co-occurrence matrix probability and statistical feature column vector x _{4,j of superpixel Y j ;} (3k) According to the order of the superpixels in the superpixel set Y, arrange the gray-level co-occurrence matrix probability and statistical feature column vectors of all superpixels in sequence to form the gray-level co-occurrence matrix statistical feature set X ₄ =(x _4,1 ,x _4,2 ,…,x _4,j ,…,x _4,n ), where, 1≤j≤n, n is the number of superpixels.