CN103353987B - A kind of superpixel segmentation method based on fuzzy theory - Google Patents

Abstract

The present invention relates to a kind of superpixel segmentation method based on fuzzy theory.Which increase the laminating degree on super-pixel border and original image border.And the super-pixel internal organizational structure after segmentation is single, uniform gray level.Step is: 1) process image, initialization cluster centre; 2) in Fuzzy C-Means Cluster Algorithm, add the coordinate distance of pixel, set up objective function; 3) subordinated-degree matrix is upgraded; 4) upgrade cluster centre, comprise gray-scale value and coordinate; 5) repetition 3-4 step, until meet end condition: the knots modification of cluster centre gray-scale value is less than a threshold value manually set or iterations is greater than certain value manually set; 6) complete preliminary super-pixel according to the subordinated-degree matrix finally obtained to divide; 7) aftertreatment, inevitably there are some isolated point sets in the super-pixel produced after above-mentioned six steps, by these point sets be adjacent and the super-pixel that similarity is the highest merge, complete final super-pixel and divide.

Description

A kind of superpixel segmentation method based on fuzzy theory

Technical field

The present invention relates to image processing field, specifically a kind of superpixel segmentation method based on fuzzy theory.

Background technology

So-called super-pixel is exactly to be condensed together by some pixel the atomic region being formed and have certain perception meaning with certain algorithm, the area grid of rigid segmentation before being used for replacing.Super-pixel can effectively utilize space constraint information, there is certain noise immunity, and strengthening, image local is conforming remains image original boundaries information simultaneously, the atomic region that super-pixel splits further comprises some characteristics of image not available for single pixel, such as shape, boundary profile information and area grayscale histogram etc., the accuracy of favourable raising image procossing, and in time complexity, super-pixel also improves a lot compared with the process of single pixel.In addition, for the gray scale uneven phenomenon that some image exists, the white matter in such as some region of brain MR image and the grey matter gray-scale value in other regions are relatively, even also low than grey matter, super-pixel is less as the grey value difference of its inside of atomic region, overall gray scale uneven phenomenon is not then present in super-pixel, effectively avoids the impact of this phenomenon on Iamge Segmentation.In recent years, super-pixel is more and more applied in Image semantic classification process.Super-pixel utilizes the similarity degree of feature between pixel and pixel to divide into groups to pixel, thus obtains the redundant information of image, reduces the complexity of successive image Processing tasks to a great extent.A large amount of research work shows, super-pixel technology has been successfully applied to repeatedly image processing tasks, as estimation of Depth, Iamge Segmentation, skeletal extraction, human body estimation and target localization etc.

Superpixel segmentation method has a lot, and method main at present has: Turbo pixel, Normalized cuts, Quick Shift, SLIC etc.The feature of Normalized cuts image segmentation result is the controllable quantity of the super-pixel produced, and shape matching is compact, and each super-pixel area is also roughly similar.But Normalized cuts algorithm speed is comparatively slow, and especially for larger picture, calculated amount is larger.The basic thought of Turbo pixel is random selecting a quantity of seeds point adopt the method for level set to expand on image, and controls the size of super-pixel block by restriction rate of growth.SLIC is that a kind of color similarity and plane of delineation space by utilizing pixel carries out cluster to pixel, thus the effective over-segmentation method generating the super-pixel of compact almost homogenization.The super-pixel that Normalized cuts, Turbo pixel and SLIC algorithm produce all is compact conformation and shape is homogeneous, therefore their semantic representation power is poor, because compact conformation makes super-pixel can not contain the overall picture of a target, and the homogeneous target of different scale that makes of shape must occur different semantic hierarchies in synonym over-segmentation.Quick-shift partitioning algorithm is a kind of pattern search dividing method based on gradient rise method.This algorithm, by constantly promoting each data point in pixel characteristic space, moves towards the nearest pixel that the gloomy density Estimation of handkerchief can be made to increase, realizes the segmentation of image.The super-pixel that Quick-shift algorithm produces is in shape be not quantitatively fixing, and the compactness of super-pixel is also poor.The most important is, these main superpixel segmentation method current are all for natural image, when adopting the medical image of the concrete ambiguity feature of these method process, effect is often unsatisfactory, the super-pixel inside of segmentation comprises medium simultaneously, is unfavorable for further image procossing.

Summary of the invention

The present invention, for overcoming above-mentioned the deficiencies in the prior art, provides a kind of superpixel segmentation method based on fuzzy theory.The method can make full use of fuzzy clustering and process the advantage in the medical image with ambiguity feature, make up the inferior position of traditional superpixel segmentation method rigid division when process has ambiguity image, improve the laminating degree on super-pixel border and original image border.And the super-pixel internal organizational structure after segmentation is single, uniform gray level.

For achieving the above object, the present invention adopts following technical scheme:

Based on a superpixel segmentation method for fuzzy theory, step is:

1) image is processed, initialization cluster centre;

2) in Fuzzy C-Means Cluster Algorithm, add the coordinate distance of pixel, set up objective function;

3) subordinated-degree matrix is upgraded;

4) upgrade cluster centre, comprise gray-scale value and coordinate;

5) repetition 3-4 step, until meet end condition: the knots modification of cluster centre gray-scale value is less than a threshold value manually set or iterations is greater than certain value manually set;

6) complete preliminary super-pixel according to the subordinated-degree matrix finally obtained to divide;

7) aftertreatment, inevitably there are some isolated point sets in the super-pixel produced after above-mentioned six steps, by these point sets be adjacent and the super-pixel that similarity is the highest merge, complete final super-pixel and divide.

In described step 1), first image is carried out rule mesh according to the super-pixel block number that will divide to format process, select the central point of each grid as initial cluster center, select the template of 3*3 around initial cluster center, using point minimum for gradient as new cluster centre.

Described step 2) in, objective function is:

$J=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{u}_{\mathrm{ik}}^p({\left|\right|y_k-v_i\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|X_k-X_i\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|Y_k-Y_i\left|\right|}^2)+\frac{\mathrm{\α}}{N_R}\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{u}_{\mathrm{ik}}^p\left(\underset{y_r\∈N_k}{\mathrm{\Σ}}{\left|\right|y_r-v_i\left|\right|}^2\right)---\left(2\right)$

Wherein, u _ikrepresent the degree of membership between a kth pixel and i cluster centre, p is the index of membership function, is used for the fog-level of control cluster result, { y _k, k=1,2 ..., the set of N} representative image gray-scale value, c is predetermined class number, { v _i, i=1,2 ..., c} is each cluster centre, and N, c are natural numbers.‖ y _k-v _i‖ ²represent the distance of pixel and cluster centre, ‖ y _r-v _i‖ ²represent the neighborhood territory pixel point of current pixel point and the distance of cluster centre, X _k, Y _krepresent the transverse and longitudinal coordinate of current pixel point respectively, X _i, Y _irepresent the transverse and longitudinal coordinate of current cluster centre point respectively, the length of side of regular grid when S represents initialization, m is the parameter of artificial setting, is used for controlling the proportion of overall distance shared by coordinate distance, N _krepresent the pixel set of current pixel point surrounding neighbors, N _rn _kradix, be set as 8, represent 8 neighborhoods around pixel, α is used for controlling the proportion of Global Information shared by neighborhood information;

Objective function comprises an implicit constraint condition g, and namely the degree of membership sum of each pixel is 1, with equation expression is:

$g=1-\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{u}_{\mathrm{ik}}---\left(3\right)$

The objective function of Problem with Some Constrained Conditions is obtained by formula (2) and (3):

F _m＝J+λg (4)

Wherein, λ is Lagrange's multiplier.

The expression-form obtaining following variable is solved with method of Lagrange multipliers to formula (4):

${u_{\mathrm{ik}}}^{*}=\frac{1}{{\mathrm{\Σ}}_{j=1}^c{((D_{\mathrm{ik}}+\frac{\mathrm{\α}}{N_R}{\mathrm{\γ}}_i)/(D_{\mathrm{jk}}+\frac{\mathrm{\α}}{N_R}{\mathrm{\γ}}_j))}^{1/(p-1)}}---\left(5\right)$

Wherein, the D in (5) formula _ik, γ _i, D _jk, γ _jbe respectively:

$D_{\mathrm{ik}}=({\left|\right|y_k-v_i\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|X_k-X_i\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|Y_k-Y_i\left|\right|}^2),$ ${\mathrm{\γ}}_i=$ $\left(\underset{y_r\∈N_k}{\mathrm{\Σ}}{\left|\right|y_r-v_i\left|\right|}^2\right),$ $D_{\mathrm{jk}}=({\left|\right|y_k-v_j\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|X_k-X_j\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|Y_k-Y_j\left|\right|}^2),$ ${\mathrm{\γ}}_j=$ $\left(\underset{y_r\∈N_k}{\mathrm{\Σ}}{\left|\right|y_r-v_j\left|\right|}^2\right).$ D _ikrepresent the distance of current pixel point and current cluster centre after with the addition of coordinate distance information, D _jkrepresent the distance of current pixel point and all cluster centres after with the addition of coordinate distance information, γ _irepresent the neighborhood territory pixel of current pixel point and the Gray homogeneity of current cluster centre, γ _jrepresent the neighborhood territory pixel of current pixel point and the Gray homogeneity of all cluster centres.I represents current cluster centre, and j represents all cluster centres.

${v_i}^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p(y_k+\frac{\mathrm{\α}}{N_R}{\mathrm{\Σ}}_{y_r\∈N_k}{y}_r)}{(1+\mathrm{\α}){\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p}---\left(6\right)$

Y _krepresent current pixel point, y _rrepresent the neighborhood territory pixel point of current pixel point.

$X_i^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p{X}_k}{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p}---\left(7\right)$

$Y_i^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p{Y}_k}{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p}---\left(8\right)$

(5) (6) (7) (8) represent the expression-form of subordinated-degree matrix, cluster centre gray-scale value, horizontal ordinate and ordinate optimum solution respectively.

In described step 6), according to the final subordinated-degree matrix obtained, obtain the category label of each pixel, the pixel with identical category mark is considered as a super-pixel, obtains preliminary division result.

In described step 7), in preliminary division result, a threshold value is set, add up the pixel number in each super-pixel, being less than the point set being considered as isolating of this threshold value, finding the super-pixel set that isolated point set with these is adjacent, finding super-pixel immediate with isolated point set by calculating average gray value, and merge with it, complete final super-pixel and divide.

The invention has the beneficial effects as follows: the present invention is directed to a kind of new superpixel segmentation method of the graphical design with ambiguity, super-pixel is generated by the coordinate distance adding pixel in Fuzzy C-Means Cluster Algorithm, the advantage of fuzzy theory in process blurred picture can be made full use of, the super-pixel welt effect generated is much better than traditional super-pixel partitioning algorithm, and super-pixel interior media is single, otherness is less.Meanwhile, invention also contemplates that the utilization of pixel surrounding neighbors information, effectively overcome the impact of noise, make the present invention have better robustness.

Accompanying drawing explanation

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

Embodiment

Below in conjunction with accompanying drawing and embodiment, the present invention will be further described.

In Fig. 1, the present invention is divided into seven steps: 1, initialization cluster centre; 2, the design of objective function; 3, subordinated-degree matrix is upgraded; 4, upgrade cluster centre, comprise gray-scale value and coordinate; 5,3-4 step is repeated, until meet end condition; 6, complete preliminary super-pixel according to the subordinated-degree matrix finally obtained to divide; 7, aftertreatment, completes the division of final super-pixel.

Idiographic flow is as follows:

1, initialization

First image is carried out rule mesh according to the super-pixel block number that will divide to format process, select the central point of each grid as initial cluster center, for avoiding the interference of noise, do a random perturbation herein, namely, select the template of 3*3 around initial cluster center, using point minimum for gradient as new cluster centre.

2, the design of objective function

The coordinate distance of pixel, when carrying out Kmeans cluster, has joined in the range formula of objective function by SLIC superpixel segmentation method, thus achieves the segmentation of super-pixel.Given this plant thought, the coordinate distance of pixel, when utilizing fuzzy clustering design object function, also adds wherein by we, the information of pixel surrounding neighbors is also taken into account meanwhile, avoids the impact of noise spot with this.The objective function of original Fuzzy C-Means Cluster Algorithm can be expressed as:

$J={\mathrm{\Σ}}_{i=1}^c{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p{\left|\right|y_k-v_i\left|\right|}^2,---\left(1\right)$

Wherein { y _k, k=1,2 ..., the set of N} representative image gray-scale value, c is predetermined class number, { v _i, i=1,2 ..., c} is each cluster centre, and N, c are natural numbers.P is the index of membership function, is used for controlling the fog-level of cluster result, ‖ y _k-v _i‖ ²represent the distance of pixel and cluster centre.The objective function that we improve is:

$J=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{u}_{\mathrm{ik}}^p({\left|\right|y_k-v_i\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|X_k-X_i\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|Y_k-Y_i\left|\right|}^2)+\frac{\mathrm{\α}}{N_R}\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{u}_{\mathrm{ik}}^p\left(\underset{y_r\∈N_k}{\mathrm{\Σ}}{\left|\right|y_r-v_i\left|\right|}^2\right)$

(2)

Wherein, u _ikrepresent the degree of membership between a kth pixel and i cluster, p is the index of membership function, is used for the fog-level of control cluster result, { y _k, k=1,2 ..., the set of N} representative image gray-scale value, c is predetermined class number, { v _i, i=1,2 ..., c} is each cluster centre, and N, c are natural numbers.‖ y _k-v _i‖ ²represent the distance of pixel and cluster centre, ‖ y _r-v _i‖ ²represent the neighborhood territory pixel point of current pixel point and the distance of cluster centre, X _k, Y _krepresent the transverse and longitudinal coordinate of current pixel point respectively, X _i, Y _irepresent the transverse and longitudinal coordinate of current cluster centre point respectively, the length of side of regular grid when S represents initialization, m is the parameter of artificial setting, is used for controlling the proportion of overall distance shared by coordinate distance, N _krepresent the pixel set of current pixel point surrounding neighbors, N _rn _kradix, be set as 8, represent 8 neighborhoods around pixel, α is used for controlling the proportion of Global Information shared by neighborhood information;

In addition, objective function comprises an implicit constraint condition g, and namely the degree of membership sum of each pixel is 1, with equation expression is:

$g=1-\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{u}_{\mathrm{ik}}---\left(3\right)$

The objective function of Problem with Some Constrained Conditions can be obtained by formula (2) and (3):

F _m＝J+λg (4)

Wherein λ is Lagrange's multiplier.

Method of Lagrange multipliers is a kind of optimized algorithm under equality constraint, and we can utilize the method to solve the optimum solution expression-form of variable in objective function.By solving to formula (4) expression-form that we can obtain following variable:

${u_{\mathrm{ik}}}^{*}=\frac{1}{{\mathrm{\Σ}}_{j=1}^c{((D_{\mathrm{ik}}+\frac{\mathrm{\α}}{N_R}{\mathrm{\γ}}_i)/(D_{\mathrm{jk}}+\frac{\mathrm{\α}}{N_R}{\mathrm{\γ}}_j))}^{1/(p-1)}}---\left(5\right)$

Wherein, the D in (5) formula _ik, γ _i, D _jk, γ _jbe respectively:

$D_{\mathrm{ik}}=({\left|\right|y_k-v_i\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|X_k-X_i\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|Y_k-Y_i\left|\right|}^2),$ ${\mathrm{\γ}}_i=$ $\left(\underset{y_r\∈N_k}{\mathrm{\Σ}}{\left|\right|y_r-v_i\left|\right|}^2\right),$ $D_{\mathrm{jk}}=({\left|\right|y_k-v_j\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|X_k-X_j\left|\right|}^2+\frac{m^2}{S^2}{\left|\right|Y_k-Y_j\left|\right|}^2),$ ${\mathrm{\γ}}_j=$ $\left(\underset{y_r\∈N_k}{\mathrm{\Σ}}{\left|\right|y_r-v_j\left|\right|}^2\right).$ D _ikrepresent the distance of current pixel point and current cluster centre after with the addition of coordinate distance information, D _jkrepresent the distance of current pixel point and all cluster centres after with the addition of coordinate distance information, γ _irepresent the neighborhood territory pixel of current pixel point and the Gray homogeneity of current cluster centre, γ _jrepresent the neighborhood territory pixel of current pixel point and the Gray homogeneity of all cluster centres.I represents current cluster centre, and j represents all cluster centres.

${v_i}^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p(y_k+\frac{\mathrm{\α}}{N_R}{\mathrm{\Σ}}_{y_r\∈N_k}{y}_r)}{(1+\mathrm{\α}){\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p}---\left(6\right)$

Y _krepresent current pixel point, y _rrepresent the neighborhood territory pixel point of current pixel point.

$X_i^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p{X}_k}{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p}---\left(7\right)$

$Y_i^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p{Y}_k}{{\mathrm{\Σ}}_{k=1}^N{u}_{\mathrm{ik}}^p}$

(8)

(5) (6) (7) (8) represent the expression-form of subordinated-degree matrix, cluster centre gray-scale value, horizontal ordinate and ordinate optimum solution respectively.

The iteration that in step below, we carry out correlated variables by these expression formulas upgrades.

3, subordinated-degree matrix is upgraded

Subordinated-degree matrix is calculated by formula (5).

4, cluster centre is upgraded

Gray-scale value and the coordinate figure of cluster centre is calculated by formula (6) (7) (8).

5, iteration asks optimum solution

Repeat 3-4 step, until meet end condition.End condition of the present invention is: the knots modification of cluster centre gray-scale value is less than a threshold value manually set or iterations is greater than certain value manually set.

6, primary segmentation result is obtained

According to five steps above, we can obtain final subordinated-degree matrix, and according to this matrix, we can obtain the category label of each pixel, the pixel with identical category mark is considered as a super-pixel, obtains primary segmentation result.

7, net result is obtained through aftertreatment

On the basis obtaining primary segmentation result, a threshold value is set, add up the pixel number in each super-pixel, be less than the point set being considered as isolating of this threshold value, find the super-pixel set that isolated point set with these is adjacent, find super-pixel immediate with isolated point set by calculating average gray value, and merge with it, complete final super-pixel and divide.

Claims (4)

1. based on a superpixel segmentation method for fuzzy theory, it is characterized in that, step is:

1) image is processed, initialization cluster centre;

2) in Fuzzy C-Means Cluster Algorithm, add the coordinate distance of pixel, set up objective function;

3) subordinated-degree matrix is upgraded;

4) upgrade cluster centre, comprise gray-scale value and coordinate;

5) step 3 is repeated)-step 4), until meet end condition: the knots modification of cluster centre gray-scale value is less than a threshold value manually set or iterations is greater than certain value manually set;

6) complete preliminary super-pixel according to the subordinated-degree matrix finally obtained to divide;

7) aftertreatment, inevitably there are some isolated point sets in the super-pixel produced after above-mentioned six steps, by these point sets be adjacent and the super-pixel that similarity is the highest merge, complete final super-pixel and divide;

Described step 2) in, objective function is:

$J=\underset{i=1}{\overset{c}{\Σ}}\underset{k=1}{\overset{N}{\Σ}}{u}_{ik}^p\left({||y_k-v_i||}^2+\frac{m^2}{S^2}{||X_k-X_i||}^2+\frac{m^2}{S^2}{||Y_k-Y_i||}^2\right)+\frac{\mathrm{\α}}{N_R}\underset{i=1}{\overset{c}{\Σ}}\underset{k=1}{\overset{N}{\Σ}}{u}_{ik}^p\left(\underset{y_r\∈N_k}{\Σ}{||y_r-v_i||}^2\right)---\left(2\right)$

Wherein, u _ikrepresent the degree of membership between a kth pixel and i cluster, p is the index of membership function, is used for the fog-level of control cluster result, { y _k, k=1,2 ..., the set of N} representative image gray-scale value, c is predetermined class number, { v _i, i=1,2 ..., c} is each cluster centre, and N, c are natural numbers; || y _k-v _i|| ²represent the distance of pixel and cluster centre, || y _r-v _i|| ²represent the neighborhood territory pixel point of current pixel point and the distance of cluster centre, X _k, Y _krepresent the transverse and longitudinal coordinate of current pixel point respectively, X _i, Y _irepresent the transverse and longitudinal coordinate of current cluster centre point respectively, the length of side of regular grid when S represents initialization, m is the parameter of artificial setting, is used for controlling the proportion of overall distance shared by coordinate distance, N _krepresent the pixel set of current pixel point surrounding neighbors, N _rn _kradix, be set as 8, represent 8 neighborhoods around pixel, α is used for controlling the proportion of Global Information shared by neighborhood information;

Objective function comprises an implicit constraint condition g, and namely the degree of membership sum of each pixel is 1, with equation expression is:

$g=1-\underset{i=1}{\overset{c}{\Σ}}{u}_{ik}---\left(3\right)$

The objective function of Problem with Some Constrained Conditions is obtained by formula (2) and (3):

F _m＝J+λg (4)

Wherein, λ is Lagrange's multiplier;

The expression-form obtaining following variable is solved with method of Lagrange multipliers to formula (4):

${u_{ik}}^{*}=\frac{1}{{\mathrm{\Σ}}_{j=1}^c{\left(\right(D_{ik}+\frac{\mathrm{\α}}{N_R}{\mathrm{\γ}}_i)/(D_{jk}+\frac{\mathrm{\α}}{N_R}{\mathrm{\γ}}_j\left)\right)}^{1/(p-1)}}---\left(5\right)$

Wherein, the D in (5) formula _ik, γ _i, D _jk, γ _jbe respectively:

$\begin{array}{c}{D}_{ik}=\left({||y_k-v_i||}^2+\frac{m^2}{S^2}{||X_k-X_i||}^2+\frac{m^2}{S^2}{||Y_k-Y_i||}^2\right),{\mathrm{\γ}}_i=\\ \left(\underset{y_r\∈N_k}{\Σ}{||y_r-v_i||}^2\right),D_{jk}=\left({||y_k-v_j||}^2+\frac{m^2}{S^2}{||X_k-X_j||}^2+\frac{m^2}{S^2}{||Y_k-Y_j||}^2\right),{\mathrm{\γ}}_j=\end{array}$ $\left(\underset{y_r\∈N_k}{\Σ}{||y_r-v_j||}^2\right);$ D _ikrepresent the distance of current pixel point and current cluster centre after with the addition of coordinate distance information, D _jkrepresent the distance of current pixel point and all cluster centres after with the addition of coordinate distance information, γ _irepresent the neighborhood territory pixel of current pixel point and the Gray homogeneity of current cluster centre, γ _jrepresent the neighborhood territory pixel of current pixel point and the Gray homogeneity of all cluster centres; I represents current cluster centre, and j represents all cluster centres;

${v_i}^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{ik}^p(y_k+\frac{\mathrm{\α}}{N_R}{\mathrm{\Σ}}_{y_r\∈N_k}{y}_r)}{(1+\mathrm{\α}){\mathrm{\Σ}}_{k=1}^N{u}_{ik}^p}---\left(6\right)$

Y _krepresent current pixel point, y _rrepresent the neighborhood territory pixel point of current pixel point;

$X_i^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{ik}^p{X}_k}{{\mathrm{\Σ}}_{k=1}^N{u}_{ik}^p}---\left(7\right)$

$Y_i^{*}=\frac{{\mathrm{\Σ}}_{k=1}^N{u}_{ik}^p{Y}_k}{{\mathrm{\Σ}}_{k=1}^N{u}_{ik}^p}---\left(8\right)$

(5) (6) (7) (8) represent the expression-form of subordinated-degree matrix, cluster centre gray-scale value, horizontal ordinate and ordinate optimum solution respectively.

2. as claimed in claim 1 based on the superpixel segmentation method of fuzzy theory, it is characterized in that, described step 1) in, first image is carried out rule mesh according to the super-pixel block number that will divide to format process, select the central point of each grid as initial cluster center, select the template of 3*3 around initial cluster center, using point minimum for gradient as new cluster centre.

3. as claimed in claim 1 based on the superpixel segmentation method of fuzzy theory, it is characterized in that, described step 6) in, according to the final subordinated-degree matrix obtained, obtain the category label of each pixel, the pixel with identical category mark is considered as a super-pixel, obtains primary segmentation result.

4. as claimed in claim 1 based on the superpixel segmentation method of fuzzy theory, it is characterized in that, described step 7) in, in preliminary division result, a threshold value is set, add up the pixel number in each super-pixel, being less than the point set being considered as isolating of this threshold value, finding the super-pixel set that isolated point set with these is adjacent, finding super-pixel immediate with isolated point set by calculating average gray value, and merge with it, complete final super-pixel and divide.