CN104732545A - Texture image segmentation method combined with sparse neighbor propagation and rapid spectral clustering - Google Patents

Abstract

The invention discloses a texture image segmentation method combined with sparse neighbor propagation and rapid spectral clustering, and the method mainly solves the problems that in an existing texture image segmentation method, the segmentation accuracy is low and the computation complexity is high. The method comprises the steps that 1, an image to be segmented is input; 2, parameters are set; 3, a grayscale-probability square matrix is generated; 4, the number of data points is counted; 5, a sparse similarity matrix is built; 6, data representative points are selected; 7, clustering is performed on the data points; 8, the image to be segmented is marked; 9, a segmented image is output. Compared with some existing texture image segmentation technologies, the texture image segmentation method combined with sparse neighbor propagation and rapid spectral clustering can better maintain the region consistency of the texture image, the texture image with high segmentation accuracy and good segmentation stability can be obtained, and the computation complexity is low.

Description

Propagate in conjunction with sparse neighbour and the texture image segmenting method of quick spectral clustering

Technical field

The invention belongs to technical field of image processing, further relate to propagating in conjunction with sparse neighbour and the texture image segmenting method of quick spectral clustering of technical field of image segmentation.The present invention can be used for the segmentation of various texture image, to reach the object of identification and evaluating objects.

Background technology

Iamge Segmentation is one of basic problem of image procossing, is the basis realizing carrying out image target identification.Wherein, Study Of Segmentation Of Textured Images is an important branch of Digital Image Processing research, is the basis of numerous graphical analysis and machine vision applications.

In image partition method, the Study Of Segmentation Of Textured Images of feature based is completed in succession by feature extraction and these two steps of Iamge Segmentation usually.The object of Study Of Segmentation Of Textured Images is that image is divided into several mutually disjoint regions by the feature such as gray scale, color, texture, space, geometric configuration according to image, require that the inside in each region has consistent texture, and the texture difference between zones of different.In the texture image segmentation of feature based, by feature extraction, each pixel textural characteristics is represented, then divide to realize Study Of Segmentation Of Textured Images to the feature set obtained.To this, clustering ensemble technology provides a kind of approach of Study Of Segmentation Of Textured Images scheme, effectively can improve the precision of Iamge Segmentation.FCM Algorithms (FCM algorithm) is the most effective, the most widely used clustering method of one and through being usually used in Study Of Segmentation Of Textured Images.FCM algorithm, based on fuzzy theory, can describe objective world more accurately, and algorithm is simple, fast convergence rate.But this algorithm is responsive and be easily absorbed in local optimum to initial cluster center, limits the degree of accuracy of cluster result.In addition, spectral clustering is also a kind of the most frequently used clustering method, although this clustering effect is pretty good, and can solve non-convex data set problem.But this algorithm needs structure adjacent degree matrix when splitting for texture maps, and needs the proper vector calculating its corresponding Laplacian Matrix.The time complexity of this two step is respectively O (n ²) and O (n ³).This high computation complexity makes it can only solve the clustering problem of data set on a small scale.When in the face of large-scale dataset, not only need the at substantial time, and a large amount of internal memory can be taken.So when by the method process Study Of Segmentation Of Textured Images, the feature of substantial amounts can make clustering algorithm operation get up to become very consuming time.

A kind of texture image segmenting method based on immune clone multiple-objection optimization is disclosed in the patent " texture image segmenting method based on immune clone multiple-objection optimization " (number of patent application 201310182014.5, publication number CN103310441A) of Xian Electronics Science and Technology University's application.The performing step of the method is: step 1, inputs texture image to be split, and extracts its eigenmatrix G; Step 2, produces initial antibodies group V (t) and carries out initial setting; Step 3, according to eigenmatrix G and antibody population V (t), calculates cluster objective function f ₁with class object function f ₂; Step 4, antagonist group V (t) carries out immune clone operation, obtains antibody population Vc (t) after cloning; Step 5, carries out nonuniform meshes operation to antibody population Vc (t) after clone, obtains antibody population Vm (t) after non-uniform variation; Step 6, carries out population recruitment operation to antibody population Vm (t) after non-uniform variation, obtains the antibody population Vm (t+1) after upgrading; Step 7, according to the antibody population Vm (t+1) after renewal and eigenmatrix G, calculates the classification of each pixel in texture image; Step 8, exports the texture image after segmentation.Although the method better can keep the region consistency of texture image, improve segmentation precision, segmentation result is made more to meet the vision of people, but the deficiency still existed is: because immune clone algorithm belongs to evolution algorithm, when using immune clone algorithm to carry out Study Of Segmentation Of Textured Images, the poor stability of segmentation result, cannot obtain effective image segmentation result.In addition, population is cloned, make a variation and the evolutionary generation of population easily increases the time of interative computation and is absorbed in local extremum, limits the degree of accuracy of cluster result.

Patented technology " a kind of dividing method of the texture image " (number of patent application 201210259652.8 that Shenzhen Graduate School of Tsinghua University has, publication number CN102819840A, Authorization Notice No. 102819840B) in disclose a kind of dividing method of texture image.The concrete steps of this patented technology are: input image to be split; The characteristic extracting different texture region in the various characteristic information token image of the local of image obtains characteristic image, reduce data volume by means such as principal component analysis (PCA)s and by mean shift algorithm, cluster is carried out to proper vector again, thus complete more reliable Study Of Segmentation Of Textured Images.Although the Iamge Segmentation reliability that this patented technology obtains relative to prior art is high, but the deficiency still existed is: this patented technology just extracts the local feature information of image to characterize the overall situation, the randomness of extraction effect is large, easily loses important information, causes segmentation effect poor.In addition, the method being reduced data volume by principal component analysis (PCA) means is easily absorbed in local optimum, and marginal information is easily lost, and limits the precision of segmentation result.

Summary of the invention

The object of the invention is to the deficiency overcoming above-mentioned prior art, propose the texture image segmenting method in conjunction with sparse neighbour's propagation and quick spectral clustering, extract the feature of texture image more accurately thus carry out Study Of Segmentation Of Textured Images sooner, improving Study Of Segmentation Of Textured Images effect.

The thinking that the present invention realizes above-mentioned purpose is: first, reads in texture image to be split, adopts generation gray scale, to the method for probability square formation, the single gray feature of image is converted to provincial characteristics; Secondly, the gray scale of acquisition is divided into the individual different region of K to probability square formation by recursive call is carried out in employing method to K mean cluster, add up the number at each number of regions strong point, and the Euclidean distance calculated in each data point and its region between other points, thus build a sparse similarity matrix; Then, adopt neighbour's propagation algorithm to act on sparse similarity matrix, obtain representative point composition and represent matrix; Finally, with quick spectral clustering, cluster is carried out to data point, export the segmentation result of image to be split.

To achieve these goals, specific implementation step of the present invention is as follows:

(1) image that input one is to be split:

Input the texture image to be split that a width size is 256 × 256 pixels;

(2) parameters:

The gray-scale value progression of the texture image to be split of input is set to 16, and maximum iteration time is set to 60;

(3) gray scale is generated to probability square formation:

(3a) with the central point of texture image to be split for initial point, set up the plane coordinate system of texture image to be split;

(3b) from texture image plane coordinate system to be split, choose arbitrarily 2 points, read the gray-scale value size corresponding to 2 chosen, 2 corresponding gray-scale values are formed a gray scale pair;

(3c) moving in parallel being used for obtaining right 2 of gray scale in step (3b) on whole coordinate plane, often mobilely once obtaining a gray scale pair, flat all-moving surface obtains 16 altogether ²plant gray scale pair;

(3d) at 256 of whole texture image coordinate plane to be split ²individual gray scale centering, statistics 16 ²plant the number that each gray scale of gray scale centering is right;

(3e) number of each gray scale on texture image coordinate plane to be split is arranged in a square formation, by this square formation normalization, obtains gray scale to probability square formation;

(4) number at statistical number strong point:

(4a) adopt K means Method, gray scale is divided into the individual different region of K to probability square formation, K >=2;

(4b) at each region recursive call K means Method, each region is divided into less region, until the data amount check in all regions is less than or equal to K;

(4c) add up and preserve the number at each number of regions strong point;

(5) sparse similarity matrix is built:

(5a) Euclidean distance between other points in each data point and its region is calculated;

(5b) point that the Euclidean distance of each data point is the shortest is chosen;

(5c) using the first row of all data points as sparse similarity matrix, the shortest point of the Euclidean distance of each data point is as the secondary series of sparse similarity matrix, and the Euclidean distance between two columns strong points is the 3rd row of sparse similarity matrix;

(6) data representative point is chosen:

Adopt neighbour's propagation algorithm, the representative point of element in compute sparse similarity matrix;

(7) cluster is carried out to data point:

(7a) representative point of element in sparse similarity matrix being arranged in a size is p × p matrix V, and p represents the number of the representative point of element in sparse similarity matrix, 0 < p≤2000;

(7b) degree of membership of cluster data collection according to the following formula, is calculated:

$u_{\mathrm{ij}}=\frac{\exp ({{-d}_{\mathrm{ij}}}^2/2\&times;{\mathrm{\&delta;}}^2)}{\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}\exp ({{-d}_{\mathrm{ij}}}^2/2\&times;{\mathrm{\&delta;}}^2)}$

Wherein, u _ijrepresent that cluster data concentrates a jth data to be under the jurisdiction of the degree of membership of i-th class, i=1,2 ..., 65536, j=1,2 ..., p, p represent the number of the representative point of element in sparse similarity matrix, and 0 < p≤2000, exp represents index operation, d _ijrepresent that gray scale is to the Euclidean distance in the i-th line number strong point to matrix V in probability square formation between jth line number strong point, δ represents gaussian kernel parameter, δ=100, and Σ represents sum operation;

(7c) by the angle value that is subordinate to of the cluster data collection of acquisition, to be arranged in a size be 65536 × p subordinated-degree matrix, 0 < p≤2000;

(7d) the p dimensional feature vector A of subordinated-degree matrix is calculated, 0 < p≤2000;

(7e) apply K means Method and cluster is carried out to proper vector A, obtain the cluster label of proper vector A;

(8) image to be split is marked:

Using the category label of the cluster label of proper vector as each pixel of image, obtain the category label of each pixel in texture image to be split;

(9) image after segmentation is exported:

Split texture image to be split according to category label, obtain the image after splitting and export segmentation result.

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

First, gray scale is generated to the method for probability square formation because the present invention adopts, the single gray feature of texture image is converted to provincial characteristics, directly do not use the gray-scale value of image as characteristic data set, overcome the shortcoming that prior art calculated amount when splitting texture image is large, make to present invention reduces computation complexity, improve processing speed.

Second, because the present invention adopts the method for K mean cluster being carried out to recursive call, the gray scale of acquisition is divided into the individual different region of K to probability square formation, overcoming the shortcoming of low, the easy loss image edge information of prior art accuracy of separation when carrying out Study Of Segmentation Of Textured Images, making to invention increases the degree of accuracy to Study Of Segmentation Of Textured Images.

3rd, because the present invention adopts neighbour's propagation algorithm to act on sparse similarity matrix, obtain representative point composition and represent the input of matrix as spectral clustering, overcome Spectral Clustering this shortcoming of height computation complexity when applying to Iamge Segmentation in prior art, overcome simultaneously prior art when carrying out Study Of Segmentation Of Textured Images to initial cluster center responsive and be easily absorbed in local optimum shortcoming, make to invention increases to the degree of accuracy of Study Of Segmentation Of Textured Images and obtain the better segmentation result of stability.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is analogous diagram of the present invention.

Embodiment

With reference to accompanying drawing 1, concrete steps of the present invention are as follows.

Step 1, inputs the texture image to be split that a width size is 256 × 256 pixels.

Step 2, parameters.

The gray-scale value progression of the texture image to be split of input is set to 16, and maximum iteration time is set to 60.

Step 3, generates gray scale to probability square formation.

With the central point of texture image to be split for initial point, set up the plane coordinate system of texture image to be split.

From texture image plane coordinate system to be split, choose arbitrarily 2 points, read the gray-scale value size corresponding to 2 chosen, 2 corresponding gray-scale values are formed a gray scale pair.

Moving in parallel being used for obtaining right 2 of gray scale on whole coordinate plane, often mobilely once obtaining a gray scale pair, flat all-moving surface obtains 16 altogether ²plant gray scale pair.

At 256 of whole texture image coordinate plane to be split ²individual gray scale centering, statistics 16 ²plant the number that each gray scale of gray scale centering is right.

The number of each gray scale on texture image coordinate plane to be split is arranged in a square formation, and by this square formation normalization, generate gray scale to probability square formation, the gray scale of acquisition is 256 × 256 to the size of probability square formation.

Step 4, the number at statistical number strong point.

Adopt K means Method, gray scale is divided into the individual different region of K to probability square formation, K >=2.

K means Method to gray scale to the step that probability square formation carries out Region dividing is: from the gray scale of image to be split to a Stochastic choice K element probability square formation as initial cluster center value, each cluster centre is divided into a class separately, K >=2, the gray scale calculating image to be split to all elements in probability square formation to the distance of K cluster centre value, relatively gray scale is to the distance of element each in probability square formation to K cluster centre value, corresponding element is given by cluster centre value category label corresponding for each element minimum value, obtain the category label of gray scale to element each in probability square formation, calculate gray scale to the mean value of dvielement every in probability square formation, obtain new cluster centre value, whether newer cluster centre value is identical with former cluster centre value, if new cluster centre value is different from former cluster centre value, then continue iteration, gray scale is re-started category division to all elements in probability square formation according to the distance with new cluster centre value, until reach maximum iteration time, export cluster result, if new cluster centre value is identical with former cluster centre value, then export cluster result.

Adopt Euclidean distance to calculate gray scale to the distance of element each in probability square formation to K cluster centre value in K means Method, Euclidean distance calculates according to the following formula:

d(x,y)＝||x-y||。

Wherein, d (x, y) represents Euclidean distance, x and y represents that gray scale is to the element of two in probability square formation respectively.

At each region recursive call K means Method, each region is divided into less region, until the data amount check in all regions is less than or equal to K, adds up and preserve the number at each number of regions strong point.

Step 5, builds sparse similarity matrix.

Calculate the Euclidean distance between other points in each data point and its region, choose the point that the Euclidean distance of each data point is the shortest.

Using the first row of all data points as sparse similarity matrix, the shortest point of the Euclidean distance of each data point is as the secondary series of sparse similarity matrix, and the Euclidean distance between two columns strong points is the 3rd row of sparse similarity matrix.

The present invention to the specific embodiment building sparse similarity matrix is: the neighbour of tentation data point 1 is 2,3,4, and the neighbour of data point 2 is 3,5,6, then sparse similarity matrix can be expressed as:

$\left(\begin{array}{ccc}1& 2& d(1,2)\\ 1& 3& d\left(\mathrm{1,3}\right)\\ 1& 4& d\left(\mathrm{1,4}\right)\\ 2& 3& d\left(\mathrm{2,3}\right)\\ 2& 5& d\left(\mathrm{2,5}\right)\\ .& .& .\\ .& .& .\\ .& .& .\\ x& y& d(x,y)\\ .& .& .\\ .& .& .\\ .& .& .\end{array}\right).$

Wherein, d (x, y) represents Euclidean distance, x and y represents that gray scale is to the element of two in probability square formation respectively.

Step 6, chooses data representative point.

Adopt neighbour's propagation algorithm, the representative point of element in compute sparse similarity matrix.

Neighbour's propagation algorithm concrete steps are as follows:

1st step, is initialized as 0 by the Attraction Degree in sparse similarity matrix between element and degree of membership, and iterations t is initialized as 1;

2nd step, according to the following formula, in compute sparse similarity matrix, data point j is to the Attraction Degree of data point i:

$r(i,j)=s(i,j)-\underset{k\&NotEqual;j,t}{\max }\{a(i,k)+s(i,k)\}.$

Wherein, r (i, j) represent that in sparse similarity matrix, data point j is to the Attraction Degree of data point i, i and j to represent in sparse similarity matrix the label of i-th and a jth element respectively, i=1, 2, 65536, j=1, 2, 65536, s (i, j) the similarity size in sparse similarity matrix between data point i and data point j is represented, max represents and gets maxima operation, k represents the label of a kth element in sparse similarity matrix, k=1, 2, 65536, t represents current iterations, t≤60, a (i, k) represent that in sparse similarity matrix, data point i is to the degree of membership of data point k, s (i, k) the similarity size in sparse similarity matrix between data point i and data point k is represented,

3rd step, according to the following formula, in compute sparse similarity matrix, data point i is to the degree of membership of data point j:

$\left\{\begin{array}{c}a(i,j)=\min \{0,r(j,j)+\underset{k\&NotEqual;\{i,j\}.t}{\mathrm{\&Sigma;}}\max \{0,r(k,j)\left\}\right\},i\&NotEqual;j\\ a(i,j)=\underset{k\&NotEqual;j.t}{\mathrm{\&Sigma;}}\max \{0,r(k,j)\},i=j\end{array}\right..$

Wherein, a (i, j) represent that in sparse similarity matrix, data point i is to the degree of membership of data point j, i and j to represent in sparse similarity matrix the label of i-th and a jth element respectively, i=1, 2, 65536, j=1, 2, 65536, min represents that getting minimum value operates, r (j, j) represent that in sparse similarity matrix, data point j is to the Attraction Degree of self, Σ represents sum operation, t represents current iterations, t≤60, max represents and gets maxima operation, r (k, j) represent that in sparse similarity matrix, data point k is to the Attraction Degree of data point j, k represents the label of a kth element in sparse similarity matrix, k=1, 2, 65536,

4th step, according to the following formula, to upgrade in sparse similarity matrix data point j to the Attraction Degree of data point i:

R(i,j)＝λ×r'(i,j)+(1-λ)×r(i,j)。

Wherein, R (i, j) represent upgrade after data point j is to the Attraction Degree of data point i in sparse similarity matrix, i and j to represent in sparse similarity matrix the label of i-th and a jth element respectively, i=1,2,, 65536, j=1,2,, 65536, λ represents convergence coefficient, λ ∈ (0,1), r'(i, j) represent current iteration before in sparse similarity matrix data point j to the Attraction Degree of data point i, r (i, j) be the 2nd step obtain sparse similarity matrix in data point j to the Attraction Degree of data point i;

5th step, according to the following formula, to upgrade in sparse similarity matrix data point i to the degree of membership of data point j:

A(i,j)＝λ×a'(i,j)+(1-λ)×a(i,j)。

Wherein, A (i, j) represent upgrade after data point i is to the degree of membership of data point j in sparse similarity matrix, i and j to represent in sparse similarity matrix the label of i-th and a jth element respectively, i=1,2,, 65536, j=1,2,, 65536, λ represents convergence coefficient, λ ∈ (0,1), a'(i, j) represent current iteration before in sparse similarity matrix data point i to the degree of membership of data point j, a (i, j) be the 3rd step obtain sparse similarity matrix in data point i to the degree of membership of data point j;

6th step, judges whether current iterations reaches the maximum iteration time 60 set by initialization, if so, performs the 7th step, otherwise, perform the 2nd step;

7th step, according to the following formula, the representative point of element in compute sparse similarity matrix:

i ^*＝argmax((A(i,j)+R(i,j))。

Wherein, i ^*represent the representative point of data point i in sparse similarity matrix, i represents the label of i-th element in sparse similarity matrix, i=1, 2, 65536, max represents and gets maxima operation, arg () represents the operation of getting value element correspondence label in sparse similarity matrix, A (i, j) represent that in the rear sparse similarity matrix of iteration renewal, data point i is to the degree of membership of data point j, j represents the label of a jth element in sparse similarity matrix, j=1, 2, 65536, R (i, j) represent that in the rear sparse similarity matrix of iteration renewal, data point j is to the Attraction Degree of data point i.

Neighbour's propagation algorithm does not need to specify clusters number, but is automatically determined by the size of bias, in embodiments of the present invention, bias is set as the median of sparse similarity matrix.

Step 7, carries out cluster to data point.

The representative point of acquisition being arranged in a size is p × p matrix V, and p represents the number of representative point, 0 < p≤2000.According to the following formula, what calculate cluster data collection is subordinate to angle value:

$u_{\mathrm{ij}}=\frac{\exp ({{-d}_{\mathrm{ij}}}^2/2\&times;{\mathrm{\&delta;}}^2)}{\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}\exp ({{-d}_{\mathrm{ij}}}^2/2\&times;{\mathrm{\&delta;}}^2)}.$

Wherein, u _ijrepresent that cluster data concentrates a jth data to be under the jurisdiction of the degree of membership of i-th class, i=1,2 ..., 65536, j=1,2 ..., p, p represent the number of the representative point of element in sparse similarity matrix, and 0 < p≤2000, exp represents index operation, d _ijrepresent that gray scale is to the Euclidean distance in the i-th line number strong point to matrix V in probability square formation between jth line number strong point, δ represents gaussian kernel parameter, δ=100, and Σ represents sum operation.

By the angle value that is subordinate to of the cluster data collection of acquisition, to be arranged in a size be 65536 × p subordinated-degree matrix, 0 < p≤2000, calculate the p dimensional feature vector A of subordinated-degree matrix, 0 < p≤2000, application K means clustering method carries out cluster to proper vector A, obtains the cluster label of proper vector A.

Step 8, marks image to be split.

Using the category label of the cluster label of proper vector as each pixel of image, obtain the category label of each pixel in texture image to be split.

Step 9, marks image to be split.

Split texture image to be split according to category label, obtain the image after splitting and export segmentation result.

Effect of the present invention further illustrates by following emulation:

1. simulated conditions:

L-G simulation test of the present invention is configured to carry out under AMD FX (tm)-6300@3.50GHz, the hardware environment of 16.0GB RAM and computer software are configured to the software environment of Matlab R2010a at computer hardware.

2. emulate content:

Fig. 2 (a) is analogous diagram of the present invention, and wherein, Fig. 2 (a) is the texture image that width two class chosen arbitrarily is different.

Emulation 1, adopt the standard FC M texture image segmenting method of prior art texture image Fig. 2 (a) different to selected width two class to split, result is as shown in Fig. 2 (b).Emulation 2, adopt the present invention's texture image Fig. 2 (a) different to selected width two class to split, result is as shown in Fig. 2 (c).

3. analysis of simulation result:

Fig. 2 (b) is the segmentation result figure that standard FC M texture image segmenting method obtains, and in figure, segmenting edge is more coarse, has a lot of assorted point, do not have good retaining zone consistance in the white portion of segmentation result.Fig. 2 (c) is the segmentation result figure that the inventive method obtains, and in figure, segmentation result of the present invention maintains the region consistency after to Study Of Segmentation Of Textured Images preferably, and segmenting edge is more regular, and segmentation result more meets the vision of people.

Add up number normalization that in the segmentation result that the present invention and standard FC M texture image segmenting method obtain, cut-point is correct respectively, obtain these two kinds of methods to the segmentation accuracy of Study Of Segmentation Of Textured Images, result is as table 1.

The segmentation accuracy of table 1 the present invention and standard FC M texture image segmenting method

Simulation type	The present invention	Standard FC M method
			Segmentation accuracy	0.9714	0.9061

As known from Table 1, the present invention is 0.9714 higher than the segmentation accuracy 0.9061 of the FCM texture image segmenting method of standard for the segmentation accuracy of texture image to be split, and the segmentation precision that invention increases Study Of Segmentation Of Textured Images is described.

Claims (3)

1. propagate in conjunction with sparse neighbour and the texture image segmenting method of quick spectral clustering, comprise the following steps:

(1) image that input one is to be split:

Input the texture image to be split that a width size is 256 × 256 pixels;

(2) parameters:

The gray-scale value progression of the texture image to be split of input is set to 16, and maximum iteration time is set to 60;

(3) gray scale is generated to probability square formation:

(3a) with the central point of texture image to be split for initial point, set up the plane coordinate system of texture image to be split;

(3b) from texture image plane coordinate system to be split, choose arbitrarily 2 points, read the gray-scale value size corresponding to 2 chosen, 2 corresponding gray-scale values are formed a gray scale pair;

(3c) moving in parallel being used for obtaining right 2 of gray scale in step (3b) on whole coordinate plane, often mobilely once obtaining a gray scale pair, flat all-moving surface obtains 16 altogether ²plant gray scale pair;

(3d) at 256 of whole texture image coordinate plane to be split ²individual gray scale centering, statistics 16 ²plant the number that each gray scale of gray scale centering is right;

(3e) number of each gray scale on texture image coordinate plane to be split is arranged in a square formation, by this square formation normalization, obtains gray scale to probability square formation;

(4) number at statistical number strong point:

(4a) adopt K means Method, gray scale is divided into the individual different region of K to probability square formation, K >=2;

(4b) at each region recursive call K means Method, each region is divided into less region, until the data amount check in all regions is less than or equal to K;

(4c) add up and preserve the number at each number of regions strong point;

(5) sparse similarity matrix is built:

(5a) Euclidean distance between other points in each data point and its region is calculated;

(5b) point that the Euclidean distance of each data point is the shortest is chosen;

(5c) using the first row of all data points as sparse similarity matrix, the shortest point of the Euclidean distance of each data point is as the secondary series of sparse similarity matrix, and the Euclidean distance between two columns strong points is the 3rd row of sparse similarity matrix;

(6) data representative point is chosen:

Adopt neighbour's propagation algorithm, the representative point of element in compute sparse similarity matrix;

(7) cluster is carried out to data point:

(7a) representative point of element in sparse similarity matrix being arranged in a size is p × p matrix V, and p represents the number of the representative point of element in sparse similarity matrix, 0 < p≤2000;

(7b) degree of membership of cluster data collection according to the following formula, is calculated:

$u_{\mathrm{ij}}=\frac{\exp (-{d_{\mathrm{ij}}}^2/2\&times;{\mathrm{\&delta;}}^2)}{\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}\exp (-{d_{\mathrm{ij}}}^2/2\&times;{\mathrm{\&delta;}}^2)}$

Wherein, u _ijrepresent that cluster data concentrates a jth data to be under the jurisdiction of the degree of membership of i-th class, i=1,2 ..., 65536, j=1,2 ..., p, p represent the number of the representative point of element in sparse similarity matrix, and 0 < p≤2000, exp represents index operation, d _ijrepresent that gray scale is to the Euclidean distance in the i-th line number strong point to matrix V in probability square formation between jth line number strong point, δ represents gaussian kernel parameter, δ=100, and Σ represents sum operation;

(7c) by the angle value that is subordinate to of the cluster data collection of acquisition, to be arranged in a size be 65536 × p subordinated-degree matrix, 0 < p≤2000;

(7d) the p dimensional feature vector A of subordinated-degree matrix is calculated, 0 < p≤2000;

(7e) apply K means Method and cluster is carried out to proper vector A, obtain the cluster label of proper vector A;

(8) image to be split is marked:

Using the category label of the cluster label of proper vector as each pixel of image, obtain the category label of each pixel in texture image to be split;

(9) image after segmentation is exported:

Split texture image to be split according to category label, obtain the image after splitting and export segmentation result.

2. according to claim 1ly propagate in conjunction with sparse neighbour and the texture image segmenting method of quick spectral clustering, it is characterized in that, described in step (4a), step (4b), the concrete steps of K means Method are as follows:

1st step, as initial cluster center value, is divided into a class by each cluster centre, K >=2 to a Stochastic choice K element probability square formation from the gray scale of image to be split separately;

2nd step, the gray scale calculating image to be split to all elements in probability square formation to the distance of K cluster centre value;

3rd step, compares gray scale to the distance of element each in probability square formation to K cluster centre value, gives corresponding element by cluster centre value category label corresponding for each element minimum value, obtain the category label of gray scale to element each in probability square formation;

4th step, calculates gray scale to the mean value of dvielement every in probability square formation, obtains new cluster centre value;

5th step, whether newer cluster centre value is identical with former cluster centre value, if new cluster centre value is different from former cluster centre value, then continue iteration, gray scale is re-started category division to all elements in probability square formation according to the distance with new cluster centre value, until reach maximum iteration time, exports cluster result, if new cluster centre value is identical with former cluster centre value, then export cluster result.

3. according to claim 1ly to propagate in conjunction with sparse neighbour and the texture image segmenting method of quick spectral clustering, it is characterized in that, described in step (6), adopt the concrete steps of neighbour's propagation algorithm acquisition representative point as follows:

1st step, is initialized as 0 by the Attraction Degree in sparse similarity matrix between element and degree of membership, and iterations t is initialized as 1;

2nd step, according to the following formula, in compute sparse similarity matrix, data point j is to the Attraction Degree of data point i:

$r(i,j)=s(i,j)-\underset{k\&NotEqual;j.t}{\max }\{a(i,k)+s(i,k)\}$

Wherein, r (i, j) represent that in sparse similarity matrix, data point j is to the Attraction Degree of data point i, i and j to represent in sparse similarity matrix the label of i-th and a jth element respectively, i=1, 2, 65536, j=1, 2, 65536, s (i, j) the similarity size in sparse similarity matrix between data point i and data point j is represented, max represents and gets maxima operation, k represents the label of a kth element in sparse similarity matrix, k=1, 2, 65536, t represents current iterations, t≤60, a (i, k) represent that in sparse similarity matrix, data point i is to the degree of membership of data point k, s (i, k) the similarity size in sparse similarity matrix between data point i and data point k is represented,

3rd step, according to the following formula, in compute sparse similarity matrix, data point i is to the degree of membership of data point j:

$\left\{\begin{array}{c}a(i,j)=\min \{0,r(j,j)+\underset{k\&NotElement;\{i,j\}.t}{\mathrm{\&Sigma;}}\max \{0,r(k,j)\left\}\right\},i\&NotEqual;j\\ a(i,j)=\underset{k\&NotEqual;j.t}{\mathrm{\&Sigma;}}\max \{0,r(k,j)\},i=j\end{array}\right.$

Wherein, a (i, j) represent that in sparse similarity matrix, data point i is to the degree of membership of data point j, i and j to represent in sparse similarity matrix the label of i-th and a jth element respectively, i=1, 2, 65536, j=1, 2, 65536, min represents that getting minimum value operates, r (j, j) represent that in sparse similarity matrix, data point j is to the Attraction Degree of self, Σ represents sum operation, t represents current iterations, t≤60, max represents and gets maxima operation, r (k, j) represent that in sparse similarity matrix, data point k is to the Attraction Degree of data point j, k represents the label of a kth element in sparse similarity matrix, k=1, 2, 65536,

4th step, according to the following formula, to upgrade in sparse similarity matrix data point j to the Attraction Degree of data point i:

R(i,j)＝λ×r'(i,j)+(1-λ)×r(i,j)

Wherein, R (i, j) represent upgrade after data point j is to the Attraction Degree of data point i in sparse similarity matrix, i and j to represent in sparse similarity matrix the label of i-th and a jth element respectively, i=1,2,, 65536, j=1,2,, 65536, λ represents convergence coefficient, λ ∈ (0,1), r'(i, j) represent current iteration before in sparse similarity matrix data point j to the Attraction Degree of data point i, r (i, j) be the 2nd step obtain sparse similarity matrix in data point j to the Attraction Degree of data point i;

5th step, according to the following formula, to upgrade in sparse similarity matrix data point i to the degree of membership of data point j:

A(i,j)＝λ×a'(i,j)+(1-λ)×a(i,j)

Wherein, A (i, j) represent upgrade after data point i is to the degree of membership of data point j in sparse similarity matrix, i and j to represent in sparse similarity matrix the label of i-th and a jth element respectively, i=1,2,, 65536, j=1,2,, 65536, λ represents convergence coefficient, λ ∈ (0,1), a'(i, j) represent current iteration before in sparse similarity matrix data point i to the degree of membership of data point j, a (i, j) be the 3rd step obtain sparse similarity matrix in data point i to the degree of membership of data point j;

6th step, judges whether current iterations reaches the maximum iteration time 60 set by initialization, if so, performs the 7th step, otherwise, perform the 2nd step;

7th step, according to the following formula, the representative point of element in compute sparse similarity matrix:

i ^*＝argmax((A(i,j)+R(i,j))

Wherein, i ^*represent the representative point of data point i in sparse similarity matrix, i represents the label of i-th element in sparse similarity matrix, i=1, 2, 65536, max represents and gets maxima operation, arg () represents the operation of getting value element correspondence label in sparse similarity matrix, A (i, j) represent that in the rear sparse similarity matrix of iteration renewal, data point i is to the degree of membership of data point j, j represents the label of a jth element in sparse similarity matrix, j=1, 2, 65536, R (i, j) represent that in the rear sparse similarity matrix of iteration renewal, data point j is to the Attraction Degree of data point i.