CN110349160B - SAR image segmentation method based on super-pixel and fuzzy C-means clustering - Google Patents

Abstract

The invention discloses a fuzzy C-means clustering SAR image segmentation method based on superpixels and sparse representation, which comprises the steps of firstly generating superpixels, then extracting gray and textural features from each subregion in an image as basic features, and providing a sparse self-representation matrix correction processing method according to scattering characteristic differences of different surface feature types of the SAR image on the basis of a sparse representation theory so as to obtain accurate distinguishing features, and finally realizing image segmentation processing with strong robustness and high operation efficiency on speckle noise in the SAR image; because the corresponding image processing is carried out on the super-pixel level, the influence of coherent speckle noise can be weakened through the integrity of the pixel set on the basis of keeping the internal information and the boundary information of the image, and meanwhile, the extracted features are more stable through integrating the information of adjacent pixels.

Description

SAR image segmentation method based on super-pixel and fuzzy C-means clustering

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a fuzzy C-means clustering SAR image segmentation method based on superpixels and sparse representation.

Background

With the development of remote sensing technology, SAR imaging technology has entered the high resolution era. Compared with the medium-low resolution SAR, the high resolution technology enhances the capability of a sensor of the high-resolution SAR to acquire the ground feature information, and can provide more complex and fine scenes and rich scattering information. However, the field of SAR image analysis and autonomous recognition is slow. Therefore, how to efficiently analyze and process a large amount of SAR image data and effectively extract information contained therein has become a large center of gravity of current SAR technical research.

The SAR image segmentation is taken as a key step of an image preprocessing stage, and essentially the whole image is divided into a plurality of image blocks with similar characteristics inside, boundaries among the blocks are obviously distinguished and are not intersected with each other, and simultaneously, a union of all the image blocks can reconstruct an original image. The purpose of image segmentation is to distinguish between regions of interest and regions of no interest, so that subsequent analytical studies can be performed only on the regions of interest that have been cut out. To obtain such pixel clusters that can be clearly distinguished and have consistent internal features, the following three general categories of algorithms are available:

(1) a threshold-based image segmentation algorithm. This method is often used in the case of double-segmented regions, and the target region is clearly distinguished from the background irrelevant region on some feature. And classifying each pixel point by defining a threshold value, and extracting a target area. Of course, this approach can be extended to multiple partitions. The most obvious defect of the threshold segmentation method is that the threshold segmentation method is greatly interfered by noise, and although the influence of the noise is considered and the suppression processing is carried out in the processing process, the segmentation precision in a noise serious area is still not ideal.

(2) Image segmentation algorithms based on edge detection. The ultimate goal of image segmentation is to form the image into several non-interfering sub-regions, so it is important to accurately locate the edges and boundaries of each sub-region. The basic idea is to detect feature points or edge points in the image and to partition the boundaries of different sub-regions based on the determination of an infinite number of points. The key of image segmentation based on the edge detection method is to identify complete and accurate sub-region contours, but this is an ideal case, and the contours of each region in most edge detection results are intermittent and incompletely connected "dotted lines", so that steps of edge completion and false edge removal are generally required. In addition, the algorithm has more processing details, and is difficult to detect reliable edges when images with more complex changes are changed, so that the accuracy of image segmentation is reduced.

(3) Region-based image segmentation algorithms. The method is mainly used for completing segmentation through consistency of internal features of the region. The method can be roughly classified into a region growing and merging method, a random field method, and a clustering method. The clustering-based segmentation algorithm extracts the correlation among different pixels, and the feature correlation is maximized by iteration to complete segmentation, thereby greatly improving the heteroscedasticity in the field of image segmentation. Fuzzy partition knowledge is combined with a clustering algorithm, and an intra-class weighted error square sum is formed by adopting a membership square weighting method, so that trivial solution is avoided. And introducing a fuzzy index m, iterating for a plurality of times, and performing optimal c-type division on the sample points by minimizing the target function, which is the well-known FCM algorithm. Also, due to their superiority, many researchers have made many studies and improvements. Such as modified FCM _ S1, FCM _ S2, EnFCM, FGFCM, NS _ FCM, and QICA _ FCM, among others.

Obviously, the most important step is to construct an objective function, or to find the correlation of features inside the image, as it were. Unlike the ordinary optical image, the SAR image contains complex speckle noise due to a unique radar imaging mechanism, and the influence of the multiplicative noise is considered in the segmentation.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides a fuzzy C-means clustering SAR image segmentation method based on superpixels and sparse representation.

In order to achieve the purpose, the invention provides a fuzzy C-means clustering SAR image segmentation method based on superpixels and sparse representation, which is characterized by comprising the following steps of:

(1) obtaining SAR image

Acquiring a single-polarized Synthetic Aperture Radar (SAR) image containing appropriate and abundant ground object types in an area to be observed;

(2) generating superpixels using PILS algorithm

(2.1) initializing the clustering center

Setting N pixel points in total of the SAR image, initializing k clustering centers, wherein each clustering center is a superpixel, and the distance between adjacent clustering centers is

Each cluster center is represented by a three-dimensional feature as: [ i, x, y ], wherein i represents a pixel intensity value of a cluster center, and x and y represent abscissa and ordinate values of the cluster center;

(2.2) Cluster center relocation

Taking a certain clustering center as a center, finding a position with the minimum gradient in an n multiplied by n neighborhood, and relocating the clustering center to the position;

(2.3) image clustering

Establishing a search domain S multiplied by S by taking a certain pixel point as a center; traversing the whole SAR image, determining a clustering center of the pixel point closest to the searching domain S multiplied by S, and dividing the pixel point into the same class of the clustering center;

(2.4) iterative update

After the category attributes of all the pixel points are distributed, calculating the mean value of the three-dimensional characteristics [ i, x, y ] of the clustering centers of all the categories, updating the clustering centers of the categories according to the mean value, and iterating until convergence;

(3) superpixel feature extraction

Calculating a gray level co-occurrence matrix GLCM in a search domain S multiplied by S to extract the texture characteristics of the super pixels, wherein the method comprises the following steps: angular second moment, contrast, correlation, inverse difference moment, variance, sum mean, sum entropy, sum variance, entropy, difference variance, correlation information measure, and maximum correlation coefficient;

extracting a gray histogram in the super pixel as a super pixel gray characteristic;

forming the super-pixel texture feature and the gray feature into an image basic feature set X (X)₁,x₂,…,x_n)，x_nRepresents the nth feature;

(4) creating a sparse self-representation model

(4.1), initializing Z-G-psi-0, wherein Z is a sparse self-expression matrix, G is a diagonal matrix, and psi is a Lagrange multiplier matrix; initializing the current iteration time t as 1;

(4.2) judging whether the t iteration is followed by the I Z^(t)-G^(t)||_∞Eta, eta isConverging a threshold, if so, entering the step (4.3), otherwise, outputting the t-th iteration to obtain a sparse self-expression matrix Z;

(4.3) column by column update G^(t)：

Wherein J is 1,2, …, J, g_j＝(g_1j,g_2j,…,g_(j-1)j,0,…,g_Jj) Mu is a penalty factor, z_jFor sparse self-representation of matrix Z^(t)J column of (1)_jIs a matrix Ψ^(t)Column j of (1); g is prepared from_jjSet to 0 to satisfy the constraint diag (G)^(t)) 0, so when updating column by column, G is encountered^(t)When the diagonal element is not updated;

updating Z^(t)：

Wherein, X is an image basic feature set, I is a unit matrix, and gamma is a setting parameter;

for updated Z^(t)And (3) correcting:

to Z^(t)Element z in (1)_ijThe following formula is used for processing to obtain the element z_ijIs estimated value of

Forming matrix by the estimated values of all elements

Go through

Calculating the backscattering mean value of each super pixel set, and enabling the mean value difference of backscattering values of the super pixels i and j to be delta sigma_ijObtaining corrected Z^(t)The elements of (A) are:

updating Ψ^(t)：

Ψ^(t)＝Ψ^(t-1)+μ(Z^(t)-G^(t))

Updating the penalty coefficient mu:

μ＝min(εμ,10¹⁰)

wherein epsilon is a constant;

(4.4) after the step (4.3) is finished, outputting a sparse self-expression matrix Z^(t)Adding 1 to the iteration times t, and returning to the step (4.2);

(5) image segmentation

(5.1) defining an objective function J of the SSR _ FCM algorithm_m：

Wherein m represents a blur index, v_k、

Respectively representing the k class centers of the matrixes X and Z, and beta is an objective function J for controlling sparse discriminant features_mWeight value of u_kiRepresenting the membership value of the superpixel i to the kth class for the introduced fuzzy membership parameter;

(5.2) for the objective function J_mCarrying out minimization:

where ξ represents the differential of the lagrange multiplier;

(5.3) solving the minimized objective function J_mObtaining:

(5.4) substitution constraint

The following can be solved:

thereby obtaining updated u_ki：

(5.5) use of the updated u_kiConstructing a membership matrix U^(k)Then, it is judged whether max (U) is full^(k)-U^(k-1))＜η^*，η^*Is a preset threshold value, if the preset threshold value is met, the step (5.6) is carried out; otherwise, enabling k to be added by 1, and then returning to the step (5.1);

(5.6) according to the membership value u_kiAnd clustering the images so as to realize image segmentation.

The invention aims to realize the following steps:

the invention relates to a fuzzy C-means clustering SAR image segmentation method based on superpixels and sparse representation, which comprises the steps of firstly generating superpixels, then extracting gray and texture characteristics of each subregion in an image as basic characteristics, and providing a sparse self-representation matrix correction processing method according to scattering characteristic differences of different surface feature types of the SAR image on the basis of a sparse representation theory so as to obtain accurate discrimination characteristics, and finally realizing image segmentation processing which has stronger robustness and high operation efficiency on speckle noise in the SAR image; because the corresponding image processing is carried out on the super-pixel level, the influence of coherent speckle noise can be weakened through the integrity of the pixel set on the basis of keeping the internal information and the boundary information of the image, and meanwhile, the extracted features are more stable through integrating the information of adjacent pixels.

Drawings

FIG. 1 is a flow chart of a fuzzy C-means clustering SAR image segmentation method based on superpixels and sparse representation according to the present invention;

FIG. 2 is a raw SAR image;

FIG. 3 is a superpixel generation diagram;

FIG. 4 is a comparison algorithm EnFCM algorithm segmentation result;

FIG. 5 is a graph of the segmentation results of the FLICM algorithm;

fig. 6 is a diagram of a segmentation result of a blurred C-means SAR image based on superpixels and sparse representation.

Detailed Description

The following description of the embodiments of the present invention is provided in order to better understand the present invention for those skilled in the art with reference to the accompanying drawings. It is to be expressly noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the subject matter of the present invention.

Examples

FIG. 1 is a flow chart of a fuzzy C-means clustering SAR image segmentation method based on superpixels and sparse representation according to the present invention;

in this embodiment, as shown in fig. 1, the method for segmenting a fuzzy C-means clustering SAR image based on superpixels and sparse representation according to the present invention includes the following steps:

s1, acquiring SAR image

Acquiring a single-polarized Synthetic Aperture Radar (SAR) image containing appropriate and abundant ground object types in an area to be observed;

in this embodiment, an SAR image of a certain area with an azimuth direction × a distance direction of 0.167 × 0.455m, which is acquired by an X-band TerraSAR in 5 months of 2014, is selected, and the size of the image is 300 × 300, as shown in fig. 2.,

s2, generating superpixel by using PILS (Pixel Intensity and Location similarity) algorithm

S2.1, initializing a clustering center

Setting N pixel points in total of the SAR image, initializing k clustering centers, wherein each clustering center is a superpixel, and the distance between adjacent clustering centers is

In this embodiment, the SAR image has 90000 pixels, which are initialized to 1200 cluster centers.

Each cluster center is represented by a three-dimensional feature as: [ i, x, y ], wherein i represents a pixel intensity value of a cluster center, and x and y represent abscissa and ordinate values of the cluster center;

s2.2, cluster center relocation

Since at initialization it is possible for the cluster center to be located at the image boundary. To avoid this, a position with the minimum gradient in the neighborhood of n × n — 3 × 3 is found with a certain cluster center as a center, and the cluster center is relocated to the position, while avoiding the influence of localization to a noise point.

S2.3, clustering images

Establishing a search domain S multiplied by S which is 5 multiplied by 5 by taking a certain pixel point as a center; traversing the whole SAR image, determining a clustering center of the pixel point closest to the searching domain S multiplied by S, and dividing the pixel point into the same class of the clustering center; compared with a K-means algorithm which needs to perform global search, the search domain of the algorithm is S multiplied by S, time consumption of distance measurement is greatly reduced, and segmentation is faster.

The following detailed description of the specific process of image clustering specifically includes:

s2.3.1, the similarity measurement step in the conventional superpixel algorithm suitable for optical images cannot be used in SAR images, and the algorithm adopts a new measurement mode, namely pixel intensity ratio distance, to define the pixel correlation in SAR images. The specific formula is as follows:

where K represents the number of search field tiles, N_i(k) Representing the tile intensity, N, of an S search field centered at i_j(k) Representing the tile strength of an S search domain centered at j;

according to the formula and aiming at the influence of multiplicative noise in the SAR image on the generation of the superpixel, the PILS algorithm defines the intensity similarity measure between pixels when the superpixel is generated by using a probability density function;

wherein I represents intensity, S_I(i, j) represents the intensity similarity measure of the clustering center i with the nearest distance of the pixel point j in the search domain S multiplied by S, K represents the number of image blocks in the search domain, N_i(k) Representing the tile intensity, N, of an S search field centered at i_j(k) Representing the tile strength of an S search domain centered at j; r is_ijIs the pattern block intensity parameter;

probability density function P (r)_ij) Satisfies the following conditions:

wherein L is SAR image vision gamma (·) representing gamma function;

s2.3.2, converting the space distance measurement between pixels into pixel space similarity measurement through standard Gaussian kernel function;

wherein xy represents space, d_xy(i, j) representing the spatial distance between the clustering center i and the pixel point j in the space;

s2.3.3, finally, the similarity measurement formula can obtain the similarity measure parameter by combining the pixel intensity similarity and the spatial position similarity:

S(i,j)＝S_I(i,j)+αS_xy(i,j)

wherein α is a weighting coefficient that balances two different similarity measures;

s2.3.4 classification of each pixel point

And determining the clustering center with the highest similarity measure parameter value of each pixel point in the search domain, and dividing the pixel points into the same type of the clustering center.

S2.4, iterative update

After the category attributes of all the pixel points are distributed, calculating the mean value of the three-dimensional features [ i, x, y ] of the clustering centers of all the categories, updating the clustering centers of the categories according to the mean value, iterating until convergence, and finally generating the superpixel result as shown in fig. 3.

S3 super-pixel feature extraction

The feature extraction is used as the basis of image analysis and is also an important link before image segmentation. The extracted features are generally expressed in a vector form, and are required to have reliability, distinctiveness and independence, so that information of pixel points or pixel areas can be accurately and unmistakably reflected. In general, computing a gray level co-occurrence matrix GLCM in a search domain sxs to extract a super-pixel texture feature includes: angular second moment, contrast, correlation, inverse difference moment, variance, sum mean, sum entropy, sum variance, entropy, difference variance, correlation information measure, and maximum correlation coefficient;

extracting a gray histogram in the super pixel as a super pixel gray characteristic;

forming the super-pixel texture feature and the gray feature into an image basic feature set X (X)₁,x₂,…,x_n)，x_nRepresents the nth feature;

s4, creating sparse self-representation model

The basic idea of sparse representation is to use a linear combination of a small number of basic signals to express most or even all of the original signal. The image is embodied in the condition that the dimensionality of the characteristic vector is less and most coefficients in the vector are 0, the main structure and the property of the image are reflected, and the specific creating process is as follows:

s4.1, initializing Z-G-psi-0, wherein Z is a sparse self-expression matrix, G is a diagonal matrix, and psi is a Lagrange multiplier matrix; initializing the current iteration time t as 1;

s4.2, judging whether the t-th iteration meets the requirement of Z^(t)-G^(t)||_∞If the eta is less than eta, the eta is a convergence threshold, if the eta is satisfied, the step S4.3 is carried out, otherwise, a sparse self-expression matrix Z is obtained after the tth iteration is output;

s4.3, updating G column by column^(t)：

Wherein J is 1,2, …, J, g_j＝(g_1j,g_2j,…,g_(j-1)j,0,…,g_Jj) Mu is a penalty factor, z_jFor sparse self-representation of matrix Z^(t)J column of (1)_jIs a matrix Ψ^(t)Column j of (1); g is prepared from_jjSet to 0 to satisfy the constraint diag (G)^(t)) 0, so when updating column by column, G is encountered^(t)When the diagonal element is not updated;

updating Z^(t)：

Wherein, X is an image basic feature set, I is a unit matrix, and gamma is a setting parameter;

for updated Z^(t)And (3) correcting:

to Z^(t)Element z in (1)_ijThe following formula is used for processing to obtain the element z_ijIs estimatedEvaluating value

When the sample basic feature set X is used for solving the sparse self-expression matrix, no matter the gray histogram feature or the texture feature in the sparse self-expression matrix, the backscattering value which is the essential feature reflecting the ground feature difference in the SAR image is not directly utilized. For a high-resolution SAR image, due to different penetration and reflection capacities of different ground objects on electromagnetic pulses, the returned echo signals show different scattering characteristics. The effective use of the backscattering value can increase the accuracy of ground object resolution. It is found from literature and empirical values that the probability of two objects being of the same class is very low when the backscattering values differ by more than 10dB between the two samples. When the backscattering values of two samples are different by more than 15dB, two ground objects belong to two types with great difference. Thus forming the estimated values of all elements into a matrix

Go through

Calculating the backscattering mean value of each super pixel set, and enabling the mean value difference of backscattering values of the super pixels i and j to be delta sigma_ijObtaining corrected Z^(t)The elements of (A) are:

by combining the backscattering value of the SAR image, the difference between different ground objects can be further expanded, different features can be distinguished more obviously, and the accuracy of subsequent processing is improved. Any element z in the sparse self-representation matrix obtained by the algorithm_ijRepresenting the similarity between sample i and sample j. z is a radical of_ijThe larger the value, the similarityThe higher the similarity, the lower the similarity between two samples, and even the samples belong to two classes. The diagonal elements represent the similarity of the sample to itself, and are set to a value of 0 without participating in the calculation. The sparse self-representation matrix embodies the inline relation among samples and can be used as a pixel feature for subsequent segmentation.

Updating Ψ^(t)：

Ψ^(t)＝Ψ^(t-1)+μ(Z^(t)-G^(t))

Updating the penalty coefficient mu:

μ＝min(εμ,10¹⁰)

wherein epsilon is a constant;

s4.4, after the step S4.3 is finished, outputting a sparse self-expression matrix Z^(t)Adding 1 to the iteration times t, and returning to the step S4.2;

s5, image segmentation

Based on Fuzzy C-Means Clustering Based on Superpixel and Sparse Representation (SSR _ FCM) SAR image segmentation algorithm, two types of features with different dimensions are mainly used during image segmentation. The first type is the image self-feature, expressed as a feature matrix X ═ X₁,x₂,…,x_n) Reflecting the respective basic texture and scattering properties of each super-pixel, called the basic features. The second type is an inline relationship between superpixels obtained through a sparse model, expressed as a feature matrix Z and called sparse discriminant features. N in the two feature matrices represents the number of superpixels, but the dimensions are different.

We now describe the specific process of segmentation:

s5.1, defining an objective function J of the SSR _ FCM algorithm_m：

Wherein m represents a blur index, v_k、

Respectively representing the k class centers of the matrixes X and Z, and beta is an objective function J for controlling sparse discriminant features_mWeight value of u_kiRepresenting the membership value of the superpixel i to the kth class for the introduced fuzzy membership parameter;

s5.2, aiming at an objective function J_mCarrying out minimization:

where ξ represents the differential of the lagrange multiplier;

s5.3, solving the minimized objective function J_mObtaining:

s5.4, substitution constraint

The following can be solved:

thereby obtaining updated u_ki：

S5.5, utilizing updated u_kiConstructing a membership matrixU^(k)Then, it is judged whether max (U) is full^(k)-U^(k-1))＜η^*，η^*Is a preset threshold value, if the preset threshold value is met, the step S5.6 is carried out; otherwise, enabling k to be added by 1, and returning to the step S5.1;

s5.6, according to the membership value u_kiAnd clustering the images so as to realize image segmentation.

FIG. 4 is a segmentation result of the EnFCM algorithm, which shows that the denoising capability is weak, and the building and lawn areas cannot be accurately distinguished;

FIG. 5 is a diagram of the segmentation result of the FLICM algorithm, which shows that the algorithm can effectively distinguish the buildings and the lawns, and the lawns have good consistency with the interior of the road area, but still have a large number of error points;

FIG. 6 is a segmentation result diagram of a fuzzy C-means SAR image based on superpixels and sparse representation, and is compared with other algorithms, when the algorithm is used for segmentation, small-area buildings in the image can be well distinguished, the edge part of the building is effectively identified, edge expansion does not occur, and the definition of the boundary is guaranteed. In addition, the road and lawn segmentation effect is good, only a small number of mistaken pixel points exist, and the region consistency is best.

Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.

Claims (2)

1. A fuzzy C-means clustering SAR image segmentation method based on superpixels and sparse representation is characterized by comprising the following steps:

(1) obtaining SAR image

Acquiring a single-polarized Synthetic Aperture Radar (SAR) image containing appropriate and abundant ground object types in an area to be observed;

(2) generating superpixels using PILS algorithm

(2.1) initializing the clustering center

Setting N pixel points in total of the SAR image, initializing k clustering centers, wherein each clustering center is a superpixel, and the distance between adjacent clustering centers is

Each cluster center is represented by a three-dimensional feature as: [ i, x, y ], wherein i represents a pixel intensity value of a cluster center, and x and y represent abscissa and ordinate values of the cluster center;

(2.2) Cluster center relocation

Taking a certain clustering center as a center, finding a position with the minimum gradient in an n multiplied by n neighborhood, and relocating the clustering center to the position;

(2.3) image clustering

Establishing a search domain S multiplied by S by taking a certain pixel point as a center; traversing the whole SAR image, determining a clustering center of the pixel point closest to the searching domain S multiplied by S, and dividing the pixel point into the same class of the clustering center;

(2.4) iterative update

After the category attributes of all the pixel points are distributed, calculating the mean value of the three-dimensional characteristics [ i, x, y ] of the clustering centers of all the categories, updating the clustering centers of the categories according to the mean value, and iterating until convergence;

(3) superpixel feature extraction

Calculating a gray level co-occurrence matrix GLCM in a search domain S multiplied by S to extract the texture characteristics of the super pixels, wherein the method comprises the following steps: angular second moment, contrast, correlation, inverse difference moment, variance, sum mean, sum entropy, sum variance, entropy, difference variance, correlation information measure, and maximum correlation coefficient;

extracting a gray histogram in the super pixel as a super pixel gray characteristic;

forming the super-pixel texture feature and the gray feature into an image basic feature set X (X)₁,x₂,…,x_n)，x_nRepresents the nth feature;

(4) creating a sparse self-representation model

(4.1), initializing Z-G-psi-0, wherein Z is a sparse self-expression matrix, G is a diagonal matrix, and psi is a Lagrange multiplier matrix; initializing the current iteration time t as 1;

(4.2) judging whether the t iteration is followed by the I Z^(t)-G^(t)||_∞If yes, entering the step (4.3), otherwise, outputting the t-th iteration to obtain a sparse self-expression matrix Z;

(4.3) column by column update G^(t)：

Wherein J is 1,2, …, J, g_j＝(g_1j,g_2j,…,g_(j-1)j,0,…,g_Jj) Mu is a penalty factor, z_jFor sparse self-representation of matrix Z^(t)J column of (1)_jIs a matrix Ψ^(t)Column j of (1); g is prepared from_jjSet to 0 to satisfy the constraint diag (G)^(t)) 0, so when updating column by column, G is encountered^(t)When the diagonal element is not updated;

updating Z^(t)：

Wherein, X is an image basic feature set, I is a unit matrix, and gamma is a setting parameter;

for updated Z^(t)And (3) correcting:

to Z^(t)Element z in (1)_ijThe following formula is used for processing to obtain the element z_ijIs estimated value of

Forming matrix by the estimated values of all elements

Go through

Calculating the backscattering mean value of each super pixel set, and enabling the mean value difference of backscattering values of the super pixels i and j to be delta sigma_ijObtaining corrected Z^(t)The elements of (A) are:

updating Ψ^(t)：

Ψ^(t)＝Ψ^(t-1)+μ(Z^(t)-G^(t))

Updating the penalty coefficient mu:

μ＝min(εμ,10¹⁰)

wherein epsilon is a constant;

(4.4) after the step (4.3) is finished, outputting a sparse self-expression matrix Z^(t)Adding 1 to the iteration times t, and returning to the step (4.2);

(5) image segmentation

(5.1) defining an objective function J of the SSR _ FCM algorithm_m：

Wherein m represents a blur index, v_k、

Respectively representing the k class centers of the matrixes X and Z, and beta is an objective function J for controlling sparse discriminant features_mWeight value of u_kiRepresenting the membership value of the superpixel i to the kth class for the introduced fuzzy membership parameter;

(5.2) for the objective function J_mCarrying out minimization:

where ξ represents the differential of the lagrange multiplier;

(5.3) solving the minimized objective function J_mObtaining:

(5.4) substitution constraint

The following can be solved:

thereby obtaining updated u_ki：

(5.5) use of the updated u_kiConstructing a membership matrix U^(k)Then, it is judged whether max (U) is full^(k)-U^(k-1))＜η^*，η^*Is a preset threshold value, if the preset threshold value is met, the step (5.6) is carried out; otherwise, enabling k to be added by 1, and then returning to the step (5.1);

(5.6) according to the membership value u_kiAnd clustering the images so as to realize image segmentation.

2. The method for segmenting the fuzzy C-means clustering SAR image based on the superpixel and the sparse representation according to claim 1, wherein in the step (2.3), the specific process of image clustering is as follows:

(2.3.1) computing intensity similarity measure between pixels when superpixels are generated by using PILS algorithm

Wherein I represents intensity, S_I(i, j) represents the intensity similarity measure of the clustering center i with the nearest distance of the pixel point j in the search domain S multiplied by S, K represents the number of image blocks in the search domain, N_i(k) Representing the tile intensity, N, of an S search field centered at i_j(k) Representing the tile strength of an S search domain centered at j; r is_ijIs the pattern block intensity parameter;

probability density function P (r)_ij) Satisfies the following conditions:

wherein L is SAR image vision gamma (·) representing gamma function;

(2.3.2) computing a spatial similarity measure between pixels at superpixel generation

Wherein xy represents space, d_xy(i, j) indicates nullThe spatial distance between the middle clustering center i and the pixel point j;

(2.3.3) calculating similarity measure parameters

S(i,j)＝S_I(i,j)+αS_xy(i,j)

Wherein α is a weighting coefficient that balances two different similarity measures;

(2.3.4) classification of each pixel

And determining the clustering center with the highest similarity measure parameter value of each pixel point in the search domain, and dividing the pixel points into the same type of the clustering center.

Images