CN103839261A - SAR image segmentation method based on decomposition evolution multi-objective optimization and FCM - Google Patents

Abstract

The invention discloses an SAR image segmentation method based on decomposition evolution multi-objective optimization and FCM. The method mainly solves the problem that in the prior art of image segmentation, image segmentation precision is not high, the evaluation index is single, and the segmentation effect is not ideal. The method comprises the steps that the Gabor feature and gray level symbiotic feature of each pixel of an image are extracted, and a superpixel is obtained through rough segmentation of a watershed, superpixel features are used as data to be clustered, a clustering center is used as individual species, the species are optimized through the decomposition evolution multi-objective method, the species obtained after evolution are used as the clustering center to initialize the FCM algorithm, a new clustering center is obtained and used as new species for participating in the next evolution of the decomposition evolution multi-objective algorithm. According to the SAR image segmentation method, the better clustering center is obtained through cross adoption of the decomposition evolution multi-objective algorithm and the FCM algorithm, the defect that the FCM initial value is sensitive and falls into a local optimal solution easily is overcome, and the better image segmentation result can be obtained.

Description

A kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM

Technical field

The invention belongs to intelligent image process field, relate to Remote Sensing Image Segmentation technology, specifically a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM, for the object of cutting apart to reach target identification to remote sensing image and synthetic-aperture radar (SAR) image, can be used for remote sensing mapping, Missile Terminal Guidance, ocean resources monitoring, military surveillance, multiple fields such as mineral resource exploration, city planning, earthquake relief work.

Background technology

Along with the rise of theory on computer vision, a focus of paying close attention to becoming image understanding field cutting apart of image, image is cut apart as front subject and has been full of challenge, has attracted numerous scholars to be engaged in this area research.It is exactly the feature according to image that image is cut apart, and segments the image into some specifically, and the region of peculiar property also extracts technology and the process of interested target.Its application is very extensive, almost appears at all spectra about image processing.

Image partition method can be divided into and utilize the single goal optimized algorithm of simple target function and the multi-objective optimization algorithm that utilizes multiple goal simultaneously to optimize according to the number of the optimization aim function adopting, in actual applications, clearly solve target if known about problem, can adopt single goal optimized algorithm, but, the pixel distribution of real image is often relatively difficult to estimate and modeling, comparatively desirable mode is searched for from multiple directions exactly simultaneously, and this just inspires researchers to adopt more wide in range multi-objective optimization algorithm to improve the combination property of Solve problems.So multi-objective optimization question application in practice is more and more subject to people's attention, multi-objective optimization algorithm is cut apart to field for image becomes the focus that scholars study gradually.

There is in recent years the image Segmentation Technology that some application multi-target methods are realized, the target of multiple mutual exclusions complementation simultaneously combines, utilize more image information to reach better segmentation result and segmentation precision, for example " based on the image segmentation algorithm of immune multi-object clustering ", this algorithm is a kind of image segmentation algorithm based on immune multi-object clustering, a kind of immunization method that adds Local Search extreme value has been proposed, and clone's population scale is carried out to self-adaptationization and then use it for image and cut apart, although the method is having certain advantage aspect region consistance and edge maintenance, but the deficiency existing is, owing to having adopted too much evolution technology, increase the computation complexity of whole cutting procedure, make splitting speed slower, simultaneously, two objective functions that the method is selected are incorrect, an objective function comprises another objective function, so just can not embody the advantage part of multi-objective Algorithm, so cause segmentation result unsatisfactory.

Summary of the invention

The object of the invention is: utilize less and some multi-objective optimization algorithms high at computation complexity for the single image information that causes of above-mentioned single-object problem evaluation index, the bad deficiency that waits of image detail retention, has proposed a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM.In the present invention, extract fusion feature as band cluster data, can reach the ability in feature extraction of good performance complement, better keep image detail; Choose two complementary objective functions, just can search for optimization from multiple directions simultaneously, make algorithm at more broad range searching, to avoid algorithm to be absorbed in locally optimal solution, improve the single shortcoming of objective function in existing method; And take full advantage of the characteristic of FCM algorithm Fast Convergent, reduce the time complexity of algorithm.

Technical scheme of the present invention is: a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM, is characterized in that: comprise the following steps:

Step 1: input remote sensing images to be split;

Step 2: extract characteristics of image to be split: utilize Gabor wave filter to extract the Gabor proper vector of image, utilize algorithm of co-matrix to extract the gray scale symbiosis proper vector of image, and proper vector using these two proper vector cascades as each pixel of image to be split;

Step 3: produce data features to be clustered: image to be split is carried out to watershed divide coarse segmentation with watershed algorithm, obtain the super pixel of former figure; All pixel features that each super pixel is comprised are averaged, represent the proper vector of each super pixel of initial clustering data, with the set of the proper vector of all super pixels as data features to be clustered, the size of features is Nf × fl, wherein Nf represents the number of coarse segmentation super pixel afterwards, and fl represents the dimension of the proper vector of each super pixel;

Step 4: utilize the initial population X={x that data initialization size to be clustered is N ₁, x ₂..., x _n, each individual x _nall represent a cluster centre, also represented a segmentation result simultaneously, n=1,2 ..., N, N is initial population size;

Step 5: the target function value F that calculates respectively each individuality according to index XB and Jm _n: using the value drawing according to index XB as target function value F _nfirst aim value, using the value drawing according to index Jm as target function value F _nsecond target value:

F _n=[F1,F2]=[XB,Jm]

Step 6: initialization ideal point Z ^*;

Wherein

the 1st minimum value that up to the present objective function XB finds, the 2nd minimum value that up to the present objective function Jm finds;

Step 7: multi-objective problem F (x)=min (F1 (x), F2 (x)) is resolved into N subproblem with Chebyshev's decomposition method, and the objective function of each concrete subproblem is as follows:

$\min \mathrm{imize}{g}_j^{\mathrm{te}}\left(x\right|{\mathrm{\λ}}^j,z^{*})=\underset{1\≤i\≤m}{\max }\left\{{\mathrm{\λ}}_i^j\right|f_{\mathrm{ji}}\left(x\right)-z_i^{*}\left|\right\}$

Wherein,

current reference point, the vector that the current optimal value of each target forms, in the present invention, the value of m is 2;

represent the objective function of j subproblem;

the weights of j subproblem; x represents a population at individual, f _ji(x) value of i objective function of the individuality correspondence of j subproblem of expression, in the present invention, the value of i equals the value of m, and value is 2;

Step 8: according to each subproblem

weights λ ^j, calculate s_n neighbours subproblem Nbor (j)=(Nbor of each subproblem _j1, Nbor _j2..., Nbor _{js_n}), Nbor _jirepresent the index of i the neighbours subproblem of j subproblem, so the value of Nborji is integer; Get s_n=10; I=1,2 ..., s_n;

Step 9: by each subproblem individual P _j(t) be initialized as x _j, x _j∈ X, wherein t is iterations, t=0; And calculate individual P _j(t) corresponding target function value Ft _j;

Step 10: to each subproblem

corresponding individual P _j(t) carry out evolutional operation and obtain interim individual P _j(t+1) "

3 neighbours subproblem s of random selection in 10.1 s_n j subproblem neighbours subproblem Nbor (j), k, l, to s, k, the individual P of l neighbours subproblem _s(t), P _k(t), P _l(t) simulate two and enter interlace operation, obtain a new interim offspring individual P _j(t+1) ';

10.2 couples of interim individual P _j(t+1) ' and carry out polynomial expression mutation operation, obtain individual P _j(t+1) ";

Step 11: to the individual P obtaining _j(t+1) " carry out an iterative operation with FCM algorithm and obtain new individual P _j(t+1);

Step 12: calculate new interim individual P _j(t+1) two target function value newF _j, and according to newF _jupgrade ideal point Z ^*; By new interim individual P _jand its desired value newF (t+1) _jupgrade all s_n neighbours subproblem Nbor (j) target function value of corresponding individual and each individual correspondence respectively of j subproblem;

Step 13: judge whether current iteration number of times t meets t<Tmax, as met, perform step 13; Otherwise, make iterations t add t=t+1 one time, return to step 11, wherein Tmax is maximum iteration time, gets Tmax=20;

Step 14: select suitable individuality as cluster centre from population: by each subproblem the individual P of parent _j(t) take out, using each individual cluster centre as super pixel that will cluster, using these cluster centres as final output disaggregation; And concentrate and select one by one body as cluster centre from final output solution according to third party's index PBM;

Step 15: the classification number that obtains each pixel: the Euclidean distance that calculates the proper vector of each pixel and the cluster centre that obtains from step 15, this pixel is grouped in the classification of cluster centre of Euclidean distance minimum apart from it, obtains the classification of each pixel;

Step 16: image is cut apart in output.

The process of the extraction characteristics of image to be split described in above-mentioned steps 2 comprises as follows:

2.1 processes of utilizing Gabor wave filter to extract the medium and low frequency texture feature vector of image include: the Gabor kernel function of bidimensional can be defined as:

$g(x,y)=\frac{1}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}_x{\mathrm{\&sigma;}}_y}\exp [-\frac{1}{2}(\frac{x^2}{{\mathrm{\&sigma;}}_x^2}+\frac{y^2}{{\mathrm{\&sigma;}}_y^2})+j2\mathrm{\&pi;F}(x\cos \mathrm{\&theta;}+y\sin \mathrm{\&theta;})]$

Wherein, σ _xand σ _yrepresent that respectively oval Gaussian function is along the standard deviation in x and y direction, F is modulating frequency, and θ is the direction of Gabor kernel function.

2.2 utilize the process of algorithm of co-matrix texture feature extraction vector to include: be first certain gray level by pending image quantization, local window size can design according to object of experiment, distance between pixel is d_n, the angular separation that makes successively again two pixel lines and transverse axis is certain angle, distinguishes according to the following formula the gray level co-occurrence matrixes of the direction number of calculation requirement:

P(i,j)=#{(x ₁,y ₁),(x ₂,y ₂)∈M×N|f(x ₁,y ₁)=r,f(x ₂,y ₂)=s}

Wherein, P (i, j) be gray level co-occurrence matrixes at the locational element of coordinate (i, j), # for set { } element number, (x ₁, y ₁) and (x ₂, y ₂) be two pixel coordinates apart from equaling d_n, ∈ is the symbol that belongs in set, the size that M × N is pending image, | be the conditional code in theory of probability, r is (x ₁, y ₁) locating the gray-scale value after pixel vector quantization, s is (x ₂, y ₂) locate the gray-scale value after pixel vector quantization.

The computation process of calculating the target function value of each individuality described in above-mentioned steps 5 includes:

The formula of the 5.1 target function value F1 that obtain according to XB index and Jm index respectively and F2 is as follows:

$\mathrm{<figure-callout\; id="1"\; label="F"\; filenames="HDA0000466989330000021.png,HDA0000466989330000032.png"\; state="\{\{state\}\}">F</figure-callout>}1=\mathrm{XB}=\frac{\underset{p=1}{\overset{k}{\mathrm{\&Sigma;}}}\underset{i,j=1}{\overset{\mathrm{Nf}}{\mathrm{\&Sigma;}}}{u}_{\mathrm{pj}}^2{\left|\right|{\mathrm{features}}_j-c_p\left|\right|}^2}{\mathrm{Nf}{\min }_{i,j}\left|\right|c_i-c_j\left|\right|}$

$F2=\mathrm{Jm}=\underset{j=1}{\overset{\mathrm{Nf}}{\mathrm{\&Sigma;}}}\underset{p=1}{\overset{k}{\mathrm{\&Sigma;}}}{u}_{\mathrm{pj}}^2{\left|\right|{\mathrm{features}}_j-c_p\left|\right|}^2$

$u_{\mathrm{pj}}=\frac{1}{\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{\left(\frac{\left|\right|c_p-{\mathrm{features}}_j\left|\right|}{\left|\right|c_i-{\mathrm{features}}_j\left|\right|}\right)}^2}$

Wherein, features _jthe pixel characteristic matrix extracting, j=1,2 ..., Nf is the line number of extracting feature eigenmatrix afterwards, i.e. the number of the image block after image coarse segmentation, c _p, p=1,2 ... k is the cluster centre of image pixel, and k is the number of the cluster centre of appointment, U _{k × Nf}it is fuzzy membership matrix;

The target function value F of 5.2 each individualities _n=[F _n1, F _n2], wherein, F _n1=F1=XB, F _n2=F2=Jm.

In above-mentioned steps 11 to the individual P obtaining _j(t+1) " carry out an iterative operation with FCM algorithm and obtain new individual P _j(t+1) described implication is as follows:

11.1 individual P _j(t+1), " as cluster centre c, obtain the fuzzy membership matrix u of this cluster centre to the Euclidean distance of data features to be clustered according to this cluster centre c _ij, u _ijbe calculated as follows:

$u_{\mathrm{ij}}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\&Sigma;}}}{\left(\frac{d_{\mathrm{ij}}}{d_{\mathrm{kj}}}\right)}^{2/(m-1)}}$

U _ijbetween 0,1, c _ithe cluster centre of dimmed group of i, d _ij=|| c _i-features _j|| represent the Euclidean distance of j the data point of i cluster centre and features, and m ∈ [1, ∞), m=2 in the present invention;

11.2 according to the 11.1 fuzzy membership matrix u that obtain _ijcluster centre c ' after must being upgraded, c ' is calculated as follows:

$c_i^{\&prime;}=\frac{\underset{j=1}{\overset{\mathrm{nf}}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{\mathrm{features}}_j}{\underset{j=1}{\overset{\mathrm{nf}}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m}$

Features _jbe j data point, nf is the number of the data point that comprises of features;

11.3 c ' that obtain are as new individual P _j(t+1);

Above-mentioned steps 12 ideal point Z ^*renewal process and more the process of the individual and corresponding desired value of the parent of new neighbor subproblem comprise as follows:

12.1 ideal point Z ^*renewal process include: if

otherwise constant; If $\mathrm{new}{F}_{j2}<Z_2^{*},$ $Z_2^{*}=\mathrm{new}{F}_{j2},$ Otherwise constant;

12.2 more all individualities of new neighbor subproblem and the process of corresponding desired value thereof include: for new interim individual P _j(t+1) each neighbours subproblem, Nbor _ji∈ Nbor (j), i=1,2 ..., s_n, wherein s_n is the number of neighbours subproblem, if for all Nbor _jihave

use new interim individual P _j(t+1) i the individuality that neighbours subproblem is corresponding of alternative j subproblem and use newF _jsubstitute i the target function value that neighbours subproblem is corresponding of j subproblem

otherwise, constant.

The process that the selection according to third party's index PBM described in above-mentioned steps 14 produces optimum segmentation result is as follows:

First 14.1 obtain the value obtaining according to third party's index PBM of each cluster centre:

$\mathrm{Ic}\left(\mathrm{index}\right)=\mathrm{PBM}\left(k\right)={(\frac{1}{k}\&times;\frac{\mathrm{Ec}}{E_k}\&times;D_k)}^2,$

$E_k=\underset{p=1}{\overset{k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{u}_{\mathrm{pj}}\left|\right|{\mathrm{features}}_j-c_p\left|\right|,D_k={\max }_{i,j=1}^k\left|\right|c_i-c_j\left|\right|$

Wherein, features _j, c _p, u _pjconsistent with the description in step 2, index=1 ... N, N is the number of the current cluster centre obtaining, k is the class number that will cut apart, E _kthe super pixel that obtains of a splitting scheme and the distance sum D of its corresponding cluster centre _kbe the maximum separation between cluster centre, Ec is each features _j, j=1 ... Nf is to the distance sum of the geometric center of features, and each cluster centre obtains Ec and equates.

Document by research about cluster index performance, referring to document MaulikU., and Bandyopadhyay S.Performance evaluation of some clustering algorithms and validity indices, IEEE Trans-actions on Pattern Analysis and Machine Intelligence, vol.24, n0.12, pp:1650-1654,2002, the document points out that PBM index is that current segmentation performance is best, index is by k, E _k, and D _kacting in conjunction between three reaches the object of finding a suitable splitting scheme.

The thinking that the present invention realizes goal of the invention is: the image of input is being carried out after feature extraction and watershed segmentation acquisition cluster data, the first random initial population that produces, choose again two complementary index XB and Jm and evaluate clustering performance as objective function, by the multi-objective Algorithm MOEA based on decomposing D and two algorithms of common clustering algorithm FCM in conjunction with the individuality in initial population is carried out to iteration optimization, after meeting stopping criterion for iteration, obtain final population, from final population, select one by one body as the cluster centre of cluster data according to third party's index PBM, according to the cluster centre obtaining, cluster data is carried out cluster and is obtained the classification number of each cluster data, obtain the segmentation result of image.

The invention has the beneficial effects as follows:

First, the present invention due to cut apart at image with process in, adopt these two complementary feature extracting methods of Gabor filtering and gray level co-occurrence matrixes to extract respectively medium and low frequency texture feature vector and the high frequency texture feature vector of image, solve prior art and only utilized Gabor filtering to extract the disappearance of the image information that image feature vector causes, made the present invention to keep preferably image detail; And, after having extracted Gabor feature, utilize the gaussian filtering that window Gabor filtering more used is large to carry out smoothing processing to the image information of extracting, can keep preferably image border, improve overall segmentation precision.

The second, in the cluster process that the present invention is cut apart at image, adopt two complementary objective functions to evaluate clustering performance, overcome the single shortcoming of prior art evaluation index, make evaluation index variation of the present invention, can obtain one group of segmentation result.

The 3rd, in the cluster process that the present invention is cut apart at image, adopt the clustering algorithm based on decomposing Evolutionary multiobjective optimization and FCM, by decomposition, multi-target evolution problem being changed into single goal subproblem one by one processes, each subproblem upgrades according to the neighbours around it, can in complicated solution space, effectively search for; And can utilize the fast characteristic of FCM algorithm the convergence speed, overcome the responsive defect with being easily absorbed in locally optimal solution of FCM initial value, thereby make the algorithm after combination can have the ability of finding fast optimum solution in global scope, make the present invention can obtain region consistance and better edge retention more accurately.

Accompanying drawing explanation

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

Fig. 2 is the present invention and the segmentation result comparison diagram of prior art on the artificial two synthetic class texture images of a width;

Fig. 3 is the present invention and the segmentation result comparison diagram of prior art on the artificial three synthetic class texture images of a width;

Fig. 4 is the present invention and the segmentation result comparison diagram of prior art on the artificial four synthetic class texture images of a width;

Fig. 5 is the present invention and the segmentation result comparison diagram of prior art on the artificial five synthetic class texture images of a width;

Fig. 6 is the present invention and the segmentation result comparison diagram of prior art on the artificial two synthetic class SAR image SAR_1 of a width (size 256 × 256);

Fig. 7 is the present invention and the segmentation result comparison diagram of prior art on one the four SAR image SAR_2 of cutting apart (size 256 × 256);

Fig. 8 is the present invention and the segmentation result comparison diagram of prior art on one the three SAR image SAR_3 of cutting apart (size 512 × 512);

Fig. 9 is the present invention and the segmentation result comparison diagram of prior art on one the four SAR image SAR_4 of cutting apart (size 440 × 440);

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further detail.

Embodiment 1

The present invention has proposed a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM, belongs to technical field of image processing, further relates to a kind of dividing method of Study Of Segmentation Of Textured Images technical field.This routine emulation is to carry out under the Pentium of dominant frequency 2.3GHZ Dual_Core CPU E5200, the hardware environment of internal memory 4GB and the software environment of MATLAB R2009a.

The present invention is a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM, deficiency that the evaluation index that exists for prior art is single, computation complexity is high, details retention is bad etc., the present invention proposes a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM.In method, extract fusion feature as data to be clustered, better kept image detail; Choose two complementary objective functions, improve in existing method that objective function is single, objective function comprises the shortcomings such as image information is few.Referring to Fig. 1, the present invention is cut apart and is comprised the following steps image:

Step 1: input texture image to be split;

Step 2: extract characteristics of image to be split;

2.1 utilize Gabor wave filter to extract the Gabor proper vector of image;

2.2 utilize algorithm of co-matrix to extract the gray scale symbiosis proper vector of image;

The gray scale symbiosis proper vector cascade of 2.3 images that the Gabor proper vector proper vector and 2.2 of 2.1 images that obtain is obtained, produces the proper vector of each pixel of image to be split;

Step 3: produce data features to be clustered: image to be split is carried out to watershed divide coarse segmentation with watershed algorithm, obtain the super pixel of former figure; Each super pixel is comprised all pixel features average, represent the proper vector of each super pixel of initial clustering data, with the set of the proper vector of all super pixels as data features to be clustered, the size of features is Nf × fl, wherein Nf represents the number of coarse segmentation super pixel afterwards, and fl represents the dimension of the proper vector of each super pixel;

Step 4: utilize the initial population X={x that data initialization size to be clustered is N ₁, x ₂..., x _n, each individual x _nall represent a cluster centre, also represented a segmentation result simultaneously, n=1,2 ..., N, N is initial population size, gets N=100;

Step 5: the target function value F that calculates respectively each individuality according to index XB and Jm _n: using the value drawing according to index XB as target function value F _nfirst aim value, using the value drawing according to index Jm as target function value F _nsecond target value:

F _n=[F1,F2]=[XB,Jm]

The formula of the 5.1 target function value F1 that obtain according to XB index and Jm index respectively and F2 is as follows:

$\mathrm{<figure-callout\; id="1"\; label="F"\; filenames="HDA0000466989330000021.png,HDA0000466989330000032.png"\; state="\{\{state\}\}">F</figure-callout>}1=\mathrm{XB}=\frac{\underset{p=1}{\overset{k}{\mathrm{\&Sigma;}}}\underset{i,j=1}{\overset{\mathrm{Nf}}{\mathrm{\&Sigma;}}}{u}_{\mathrm{pj}}^2{\left|\right|{\mathrm{features}}_j-c_p\left|\right|}^2}{\mathrm{Nf}{\min }_{i,j}\left|\right|c_i-c_j\left|\right|}$

$F2=\mathrm{Jm}=\underset{j=1}{\overset{\mathrm{Nf}}{\mathrm{\&Sigma;}}}\underset{p=1}{\overset{k}{\mathrm{\&Sigma;}}}{u}_{\mathrm{pj}}^2{\left|\right|{\mathrm{features}}_j-c_p\left|\right|}^2$

$u_{\mathrm{pj}}=\frac{1}{\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{\left(\frac{\left|\right|c_p-{\mathrm{features}}_j\left|\right|}{\left|\right|c_i-{\mathrm{features}}_j\left|\right|}\right)}^2}$

Wherein, features _j, j=1,2 ..., Nf is the line number of extracting feature eigenmatrix afterwards, i.e. the number of the image block after image coarse segmentation, c _p, p=1,2 ... k is the cluster centre of image pixel, and k is the number of the cluster centre of appointment, U _{k × Nf}it is fuzzy membership matrix;

The target function value F of 5.2 each individualities _n=[F _n1, F _n2], wherein, F _n1=F1=XB, F _n2=F2=Jm;

Step 6: initialization ideal point Z ^*;

Wherein

the 1st minimum value that up to the present objective function XB finds,

the 2nd minimum value that up to the present objective function Jm finds;

Step 7: multi-objective problem F (x)=min (F1 (x), F2 (x)) is resolved into N subproblem with Chebyshev's decomposition method, and the objective function of each concrete subproblem is as follows:

$\min \mathrm{imize}{g}_j^{\mathrm{te}}\left(x\right|{\mathrm{\&lambda;}}^j,z^{*})=\underset{1\&le;i\&le;m}{\mathrm{<figure-callout\; id="1"\; label="max"\; filenames="HDA0000466989330000021.png,HDA0000466989330000032.png"\; state="\{\{state\}\}">max</figure-callout>}}\left\{{\mathrm{\&lambda;}}_i^j\right|f_{\mathrm{ji}}\left(x\right)-z_i^{*}\left|\right\}$

Wherein,

current reference point, the vector that the current optimal value of each target forms, in the present invention, the value of m is 2;

represent the objective function of j subproblem;

the weights of j subproblem;

x represents a population at individual, f _ji(x) value of i objective function of the individuality correspondence of j subproblem of expression, in the present invention, the value of i equals the value of m, and value is 2;

Step 8: according to each subproblem weights λ ^j, calculate s_n neighbours subproblem Nbor (j)=(Nbor of each subproblem _j1, Nbor _j2..., Nbor _{js_n}), Nbor _jirepresent the index of i the neighbours subproblem of j subproblem, so Nbor _jivalue be integer; Get s_n=10; I=1,2 ..., s_n;

Step 9: by each subproblem

individual P _j(t) be initialized as x _j, x _j∈ X, wherein t is iterations, t=0; And calculate individual P _j(t) corresponding target function value Ft _j;

Step 10: to each subproblem individual P _j(t) carry out evolutional operation and obtain interim individual P _j(t+1) "

3 neighbours subproblem s of random selection in 10.1 s_n j subproblem neighbours subproblem Nbor (j), k, l, to s, k, the individual P of l neighbours subproblem _s(t), P _k(t), P _l(t) carry out interlace operation, obtain a new interim offspring individual P _j(t+1) ';

10.2 couples of interim individual P _j(t+1) ' and carry out polynomial expression mutation operation, obtain individual P _j(t+1) ";

3 neighbours subproblems are selected in random selection, can expand the scope of search volume, in larger search volume, search for, and just can jump out local optimum, find better solution.

Step 11: to the individual P obtaining _j(t+1) " carry out an iterative operation with FCM algorithm and obtain new individual P _j(t+1)

11.1 individual P _j(t+1), " as cluster centre c, obtain the fuzzy membership matrix u of this cluster centre to the Euclidean distance of data features to be clustered according to this cluster centre c _ij, u _ijbe calculated as follows:

$u_{\mathrm{ij}}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\&Sigma;}}}{\left(\frac{d_{\mathrm{ij}}}{d_{\mathrm{kj}}}\right)}^{2/(m-1)}}$

U _ijbetween 0,1, c _ithe cluster centre of dimmed group of i, d _ij=|| c _i-features _j|| represent the Euclidean distance of j the data point of i cluster centre and features, and m ∈ [1, ∞), m=2 in the present invention;

11.2 according to the 11.1 fuzzy membership matrix u that obtain _ijcluster centre c ' after must being upgraded, c ' is calculated as follows:

$c_i^{\&prime;}=\frac{\underset{j=1}{\overset{\mathrm{nf}}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{\mathrm{features}}_j}{\underset{j=1}{\overset{\mathrm{nf}}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m}$

Features _jbe j data point, nf is the number of the data point that comprises of features;

11.3 c ' that obtain are as new individual P _j(t+1);

Step 12: calculate new interim individual P _j(t+1) two target function value newF _j, and according to newF _jupgrade ideal point Z ^*; By new interim individual P _jand its desired value newF (t+1) _jupgrade all s_n neighbours subproblem Nbor (j) target function value of corresponding individual and each individual correspondence respectively of j subproblem;

Step 13: judge whether current iteration number of times t meets t<Tmax, as met, perform step 13; Otherwise, make iterations t add t=t+1 one time, return to step 11, wherein Tmax is maximum iteration time, gets Tmax=20;

Step 14: select suitable individuality as cluster centre from population: by each subproblem the individual P of parent _j(t) take out, using each individual cluster centre as super pixel that will cluster, using these cluster centres as final output disaggregation; And concentrate and select one by one body as cluster centre from final output solution according to third party's index PBM;

Step 15: the classification number that obtains each pixel: the Euclidean distance that calculates the proper vector of each pixel and the cluster centre that obtains from step 15, this pixel is grouped in the classification of cluster centre of Euclidean distance minimum apart from it, obtain the classification of each pixel, and and cut apart template figure and contrast, obtain error parameter;

Step 16: image is cut apart in output.

Embodiment 2

SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM, with embodiment 1, in order to possess exploitativeness, is further described in detail as follows the present invention:

Wherein being further described in detail as follows of image characteristics extraction in step 2:

2.1.1 the process of utilizing Gabor wave filter to extract the medium and low frequency texture feature vector of image includes: the Gabor kernel function of bidimensional can be defined as:

$g(x,y)=\frac{1}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}_x{\mathrm{\&sigma;}}_y}\exp [-\frac{1}{2}(\frac{x^2}{{\mathrm{\&sigma;}}_x^2}+\frac{y^2}{{\mathrm{\&sigma;}}_y^2})+j2\mathrm{\&pi;F}(x\cos \mathrm{\&theta;}+y\sin \mathrm{\&theta;})]$

Wherein, σ _xand σ _yrepresent that respectively oval Gaussian function is along the standard deviation in x and y direction, F is modulating frequency, and θ is the direction of Gabor kernel function, because Gabor wave filter is at the conjugate symmetry of frequency domain, therefore only need be in 0-180 degree choice direction parameter θ.We are according to document ClausiD.A and DengH.Design-based texture feature fusion using Gabor filters and co-occurrence probabilities in the present invention, IEEE Transaction on Image Processing, vol.14, no.7, pp:925-936,2005. choose six centre frequency F=6.1876,4.3878,3.9135,3.6751,3.3991,2.9551, and six kernel function directions

as the parameter of wave filter.Obtain the texture feature vector of 36 dimensions of each pixel.

2.2.1 utilize the process of algorithm of co-matrix texture feature extraction vector to include: to be first 32 gray levels by pending image quantization, local window size is 9 × 9, distance between pixel is 1, the angular separation that makes successively again two pixel lines and transverse axis is 0 °, 45 °, 90 ° and 135 °, calculates respectively according to the following formula the gray level co-occurrence matrixes of four direction:

P(i,j)=#{(x ₁,y ₁),(x ₂,y ₂)∈M×N|f(x ₁,y ₁)=r,f(x ₂,y ₂)=s}

Wherein, P (i, j) be gray level co-occurrence matrixes at the locational element of coordinate (i, j), # for set { } element number, (x ₁, y ₁) and (x ₂, y ₂) be two pixel coordinates that distance equals 1, ∈ is the symbol that belongs in set, the size that M × N is pending image, | be the conditional code in theory of probability, r is (x ₁, y ₁) locating the gray-scale value after pixel vector quantization, s is (x ₂, y ₂) locate the gray-scale value after pixel vector quantization.

Choose respectively contrast, homogeney and the energy value on this matrix four direction according to gray level co-occurrence matrixes, finally obtain 12 dimension texture feature vectors of pixel.

In step 3, utilize dividing ridge method to carry out coarse segmentation to image to be split, the concrete watershed segmentation method that the present invention uses is the watershed segmentation method based on gradient.

The method can be referring to K.Haris, S.N.Efstratiadis, N.Maglaveras, and A.K.Katsaggelos, " Hybrid image segmentation using watersheds and fast region merging, " IEEE Transactions on Image Processing, Vol.7, No.12, pp.1684-1699,1998.

In the cluster process that in step 5, the present invention is cut apart at image, adopt XB index and Jm index to evaluate clustering performance as two complementary objective functions, overcome the single shortcoming of prior art evaluation index.Referring to document Bandyopadhyay S., Maulik U., and Mukhopadhyay A.Multi-objective genetic clustering for pixel classification in remote sensing imagery, IEEE Transaction on Geoscience and Remote Sensing, Vol.45, no.5, pp:1506-1511,2007.The combination of these two complementary targets makes evaluation index variation of the present invention, is more suitable for the complex information that remote sensing images comprise, and can obtain better effect.Single goal method is moved the result repeatedly obtaining, and multi-target method only need to move and once just can obtain one group of segmentation result.

In the cluster process that in step 7, the present invention is cut apart at image, adopted based on decompose Evolutionary multiobjective optimization MOEA D-algorithm, by MOEA D-algorithm multi-target evolution problem changed into single goal subproblem one by one process, each subproblem upgrades according to the neighbours around it, can in complicated solution space, effectively search for, overcome prior art and be easily absorbed in local optimum and affect the shortcoming of segmentation result, and reduced the computation complexity of the every generation of holistic approach.

Being implemented as follows of step 10.1 in step 10 in the individual evolution of each subproblem operation:

10.1.1 simulate two and enter to intersect to produce new interim individual P _j(t+1) ' process as follows:

3 neighbours subproblem s of random selection in s_n the neighbours subproblem Nbor (j) of j subproblem, k, l, to s, k, the individual P of l neighbours subproblem _s(t), P _k(t), P _l(t) simulate two and enter interlace operation, each individuality is the vector of a fl dimension, for example, and individual P _s(t) can be expressed as

the formula producing is as follows:

$P_j^i{(t+1)}^{\&prime;}=\left\{\begin{array}{cc}{P}_s^i\left(t\right)+F(P_k^i\left(t\right)-P_l^i\left(t\right)),& \mathrm{if}{\mathrm{rand}}_n\left(\mathrm{0,1}\right)<\mathrm{Cr}\\ P_j^i\left(t\right),& \mathrm{otherwise}\end{array}\right.$

Wherein Cr ∈ [0,1] is crossover probability, and F is a constant factor, is set to 0.5,

represent individual i position.

10.2.1 adopt the method for polynomial expression variation to individual P _j(t+1) ' making a variation produces new new individual P _j(t+1) ", if to make a variation be

be individual P _j(t+1) ' k (1≤k≤fl) position, its its span is [l _k, u _k] formula is as follows:

$P_j^i{(t+1)}^{\&prime;\&prime;}=\left\{\begin{array}{cc}{P}_j^i{(t+1)}^{\&prime;},& i\&NotEqual;k\\ C_k,& i=k\end{array}\right.$

Wherein △ is called as variation step-length, and its computing formula is: △=δ × (u _k-l _k), being wherein expressed as follows of δ:

$\mathrm{\&delta;}=\left\{\begin{array}{ccc}{[2u+(1-2u){(1-{\mathrm{\&delta;}}_1)}^{{\mathrm{\&eta;}}_m+1}]}^{\frac{1}{{\mathrm{\&eta;}}_m+1}-1}& \mathrm{if}& u\&le;0.5\\ 1-{[2(1-u)+2(u-0.5){(1-{\mathrm{\&delta;}}_2)}^{{\mathrm{\&eta;}}_m+1}]}^{\frac{1}{{\mathrm{\&eta;}}_m+1}-1}& \mathrm{if}& u>0.5\end{array}\right.$

δ in formula ₁=(ν _k-l _k)/(u _k-l _k), δ ₂=(u _k-ν _k)/(u _k-l _k), u is the random number in [0,1] interval, η _mit is profile exponent.

In step 14, producing optimum segmentation result detailed process includes:

First 14.1 obtain the value obtaining according to third party's index PBM of each cluster centre:

$\mathrm{Ic}\left(\mathrm{index}\right)=\mathrm{PBM}\left(k\right)={(\frac{1}{k}\&times;\frac{\mathrm{Ec}}{E_k}\&times;D_k)}^2,$

$E_k=\underset{p=1}{\overset{k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{u}_{\mathrm{pj}}\left|\right|{\mathrm{features}}_j-c_p\left|\right|,D_k={\max }_{i,j=1}^k\left|\right|c_i-c_j\left|\right|$

Wherein, features _j, c _p, u _pjconsistent with the description in claim 2, index=1 ... N, N is the number of the current cluster centre obtaining, k is the class number that will cut apart, E _kthe super pixel that obtains of a splitting scheme and the distance sum D of its corresponding cluster centre _kbe the maximum separation between cluster centre, Ec is each features _j, j=1 ... Nf is to the distance sum of the geometric center of features, and each cluster centre obtains Ec and equates;

The 14.2 Ic vectors that obtain from 14.1 steps, find the subscript en at maximum value place, concentrate and select en cluster centre from final output solution, Here it is according to PBM index cluster centre out.

Embodiment 3

Multiple-target remote sensing image dividing method based on decomposing is with embodiment 1-2, and segmentation effect of the present invention can further illustrate by following experiment:

Experiment simulation environment is: Pentium Dual_Core CPU E5200, the hardware environment of internal memory 2GB and the software environment of MATLAB R2009a of dominant frequency 2.3GHz.

In experiment also by Fuzzy C-Means Cluster Algorithm of the prior art (FCM), IMIS (merging the immune multi-object image segmentation algorithm of Gabor filtering and gray scale symbiosis complementary characteristic), NSGA-II (A fast and elitist multi-objective genetic algorithm) algorithm were also applied to respectively in cutting apart of former figure, compared with the present invention and above-mentioned three kinds of dividing methods.

Arranging of the setting of algorithm of the present invention and contrast algorithm is as follows, and the maximum iteration time of FCM is 120 times, and Fuzzy Exponential is 2, and independent operating 30 times, therefrom chooses best result.NSGA-II, IMIS and algorithm are herein all multiple-objection optimization evolution algorithms, and the setting of parameter is as following table 1:

Table 1:

Fig. 2 (a), 3 (a), 4 (a), 5 (a), are respectively the artificial synthetic former figure of texture maps two Texture Segmentation using in emulation experiment, and texture is cut apart former figure, the former figure of four Texture Segmentation and the former figure of five Texture Segmentation, image size is 256 × 256; 2 (b), 3 (b), 4 (b), 5 (b), be respectively the template of cutting apart of manually synthesizing texture maps two Texture Segmentation figure of using in emulation experiment, texture is cut apart the template of cutting apart of figure, the template of cutting apart of cutting apart template and five Texture Segmentation figure of four Texture Segmentation figure; 2 (c), 3 (c), 4 (c), 5 (c), being respectively in emulation experiment uses FCM algorithm to manually synthesizing the segmentation result of texture maps two Texture Segmentation figure, texture is cut apart the segmentation result of figure, the segmentation result of cutting apart template and five Texture Segmentation figure of four Texture Segmentation figure; 2 (d), 3 (d), 4 (d), 5 (d), being respectively in emulation experiment uses IMIS algorithm to manually synthesizing the segmentation result of texture maps two Texture Segmentation figure, texture is cut apart the segmentation result of figure, the segmentation result of cutting apart template and five Texture Segmentation figure of four Texture Segmentation figure; 2 (e), 3 (e), 4 (e), 5 (e), being respectively in emulation experiment uses NSGA-II algorithm to manually synthesizing the segmentation result of texture maps two Texture Segmentation figure, texture is cut apart the segmentation result of figure, the segmentation result of cutting apart template and five Texture Segmentation figure of four Texture Segmentation figure; 2 (f), 3 (f), 4 (f), 5 (f), be respectively in emulation experiment the segmentation result to artificial synthetic texture maps two Texture Segmentation figure with algorithm of the present invention, texture is cut apart the segmentation result of figure, and the segmentation result of cutting apart template and five Texture Segmentation figure of four Texture Segmentation figure is in table 2;

Table 2 has provided the XB value of four kinds of methods, Jm value, PBM value, cluster accuracy AR (Accuracyrate) value, and the comparison of ARI (AdjustedRandIndex) desired value, wherein the result of FCM is that isolated operation is got best result 30 times.

Best value in four algorithms of expression of overstriking mark, the analysis of experimental result, Fig. 2, 3, 4 and 5 have described respectively algorithm FCM, NSGA-II algorithm, IMIS algorithm, be respectively used to cut apart the synthetic texture maps of two classes with algorithm of the present invention, three classes are synthesized texture maps, four classes are synthesized texture maps, and the result of the synthetic texture maps of five classes, table 2 has provided the XB of these four algorithms to different texture maps segmentation results, Jm, PBM, the result of ARI evaluation index, the value of XB is the smaller the better, the index of Jm is also the smaller the better, and two indexs are difficult to reach and minimize simultaneously, the value of PBM is to be the bigger the better, the value of ARI is also the bigger the better.

Table 2:

Can draw from Fig. 2, four algorithms have all had reasonable segmentation result to two class Texture Segmentation figure, so can draw after the Fusion Features extracting respectively with Gabor and gray level co-occurrence matrixes herein, can catch preferably the information of image, greatly reduce data volume, simplify the cutting procedure of image, Fig. 2, 3 and 4 can find out, algorithm FCM, IMIS, algorithm of the present invention has all been obtained good segmentation result, but associative list 2 finds that the segmentation result accuracy of algorithm of the present invention is except two Texture Segmentation are leading by FCM, at texture figure, four texture maps, the cluster of five texture maps strives for that rate all will be higher than other algorithm, segmentation result is also all better than other algorithm, although the FCM algorithm of single goal can obtain once segmentation result preferably, but find in the process of experiment, its segmentation result is unstable, isolated operation 30 times, its average statistical is unstable, to be inferior in this respect other algorithm,

It should be noted that, as seen from Table 2, in the time that three cut apart figure, FCM has obtained minimum Jm value, but PBM, AR, ARI refers to that target value is not but best, and although algorithm of the present invention cannot be obtained the minimum value of XB and Jm simultaneously, but the index of segmentation result is but optimum, illustrate that optimum solution is not the extreme value place that is present in these two targets, but be present in their compromise positions, therefore, the model of multi-objective Algorithm can mate this demand better, the compromise position that finds single goal optimized algorithm to be not easy to find, be conducive to solving of problem, obtain being more suitable for the target solution in this problem.

Embodiment 4:

Multiple-target remote sensing image dividing method based on decomposing is with embodiment 1-2, and segmentation effect of the present invention can further illustrate by following experiment:

Experiment simulation environment is: Pentium Dual_Core CPU E5200, the hardware environment of internal memory 2GB and the software environment of MATLAB R2009a of dominant frequency 2.3GHZ.

In experiment also by Fuzzy C-Means Cluster Algorithm of the prior art (FCM), IMIS (merging the immune multi-object image segmentation algorithm of Gabor filtering and gray scale symbiosis complementary characteristic), NSGA_II (A fast and elitist multi-objective genetic algorithm) algorithm are also applied to respectively in the cutting apart of SAR image, and with this, clearly demarcated and above-mentioned three kinds of dividing methods compare.

Arranging of the setting of algorithm of the present invention and contrast algorithm is as follows, and the maximum iteration time of FCM is 120 times, and Fuzzy Exponential is 2, and independent operating 30 times, therefrom chooses best result.NSGA-II, IMIS and algorithm are herein all multiple-objection optimization evolution algorithms, and the setting of parameter is as following table 3:

Table 3:

Fig. 6 (a), 7 (a) are respectively the former figure of the artificial two synthetic class SAR image SAR_1 that use in emulation experiment, the former figure of the four SAR image SAR_2 of cutting apart, image size is 256 × 256,8 (a) are other one the three former figure of SAR image SAR_3 of cutting apart using in emulation experiment, size is 512 × 512, and 9 (a) are other one the four former figure of SAR image SAR_4 of cutting apart using in emulation experiment, and size is 440 × 440; 6 (b) are the template of cutting apart of two class SAR image SAR_1 of the artificial composite diagram used in emulation experiment; 6 (c), 7 (b), 8 (b), 9 (b) are respectively the segmentation result of algorithm FCM to SAR_1, SAR_2, SAR_3 and SAR_4; 6 (d), 7 (c), 8 (c), 9 (c) are respectively the segmentation result of IMIS algorithm to SAR_1, SAR_2, SAR_3 and SAR_4; 6 (e), 7 (d), 8 (d), 9 (d) are respectively the segmentation result of NSAG-II algorithm to SAR_1, SAR_2, SAR_3 and SAR_4; 6 (f), 7 (e), 8 (e), 9 (e) are respectively the segmentation result of algorithm of the present invention to SAR_1, SAR_2, SAR_3 and SAR_4;

Table 4:

Table 4 has provided the XB value of four kinds of methods, Jm value, PBM value, cluster accuracy AR (Accuracy rate) value, and the comparison of ARI (Adjusted Rand Index) desired value, wherein the result of FCM is that isolated operation is got best result 30 times.

Fig. 6 (a) is to the artificial two synthetic SAR figure of cutting apart, can find out, four algorithms have all been obtained good segmentation result, this illustrates that it is also feasible that our feature extracting method is applied to SAR figure, can catch preferably the textural characteristics information of SAR figure, can be further used for cutting apart.Fig. 7 is the segmentation result contrast of the true SAR figure to secondary four classes, 7 (a) comprise airfield runway, residential quarter, the Si Ge region, lawn of top-right lawn and lower left, very high to the requirement of dividing property of frontier district, especially around residential quarter, existing runway has again region, lawn.From segmentation result, IMIS algorithm is well divided runway to come, but region, two lawns does not make a distinction, residential quarter also has a lot of mixed and disorderly regions around, cause segmentation result unintelligible, FCM algorithm and NSGA-II algorithm do not have four regions separated, and the present invention makes a distinction piece region preferably, and image detail processing is better than IMIS algorithm.Fig. 8 is also the SAR figure that a width real three is cut apart, comprise farmland, shrub and river, the atural objects such as bridge can be found out FCM, IMIS and algorithm of the present invention are by all river and shrub are substantially separated, NSGA-II has but been divided into a class farmland and river, the middle part, segmentation result image the right of FCM algorithm and a part of farmland in the upper upper left corner have all been divided into river, in the segmentation result of IMIS algorithm, some is divided into river one class the farmland of middle part, image the right part, and algorithm of the present invention is compared other three algorithms, the image area of wrong point is minimum, in table 4, algorithm of the present invention obtains about the PBM value of Fig. 8 (a) also maximum simultaneously.Fig. 9 (a) is also the SAR figure that a width real four is cut apart, formed by residential block and different farmlands, cut apart difficult point and be the differentiation between division and the different croplands of residential block, from segmentation result, four algorithms are more or less the same to the local division result of residential block and farmland handing-over, but in Fig. 9 (b), farmland, image middle part segmentation result has mixed and disorderly point, Fig. 9 (c) is also that the farmland at image middle part is divided unintelligible, and top right-hand side part farmland division result is fuzzy, it is true that in Fig. 9 (d), unclassified is divided in farmland, and the farmland of image top is also divided for residential block, Fig. 9 (e) by clear farmland division out, and the interface portion in residential block and farmland is also more clearly, associative list 4, the PBM of algorithm of the present invention is to be also maximum.

In sum, prove that algorithm of the present invention cuts apart and obtained comparatively satisfied result at SAR image.The SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM that the present invention proposes, extracts fusion feature as data to be clustered, can better keep image detail; Choose two complementary objective functions, improve in existing method that objective function is single, an objective function comprises the shortcomings such as another objective function.The subproblem that the present invention is resolved into a series of Weighted Coefficients by multi-objective problem in the method based on decomposing is solved, and reduces computation complexity, has improved the precision that general image is cut apart.And take full advantage of the characteristic of FCM Fast Convergent, use based on decomposing Evolutionary Multiobjective Optimization simultaneously and overcome the responsive defect with being easily absorbed in locally optimal solution of FCM initial value, thereby make the algorithm after combination can have the ability of finding fast optimum solution in global scope, the image segmentation result that obtains being comparatively satisfied with.

Claims (6)

1. the SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM, is characterized in that: comprise the following steps:

Step 1: input remote sensing images to be split;

Step 2: extract characteristics of image to be split: utilize Gabor wave filter to extract the Gabor proper vector of image, utilize algorithm of co-matrix to extract the gray scale symbiosis proper vector of image, and proper vector after two feature vectors cascades is as the proper vector of each pixel of image to be split;

Step 3: produce data features to be clustered: image to be split is carried out to watershed divide coarse segmentation with watershed algorithm, obtain the super pixel of former figure; All pixel features that each super pixel is comprised are averaged, represent the proper vector of each super pixel of initial clustering data, with the set of the proper vector of all super pixels as data features to be clustered, the size of features is Nf × fl, wherein Nf represents the number of coarse segmentation super pixel afterwards, and fl represents the dimension of the proper vector of each super pixel;

Step 4: utilize the initial population X={x that data initialization size to be clustered is N ₁, x ₂..., x _n, each individual x _nall represent a cluster centre, also represented a segmentation result simultaneously, n=1,2 ..., N, N is initial population size;

Step 5: the target function value F that calculates respectively each individuality according to index XB and Jm _n: using the value drawing according to index XB as target function value F _nfirst aim value, using the value drawing according to index Jm as target function value F _nsecond target value:

F _n=[F1,F2]=[XB,Jm]

Step 6: initialization ideal point Z ^*;

Wherein

the 1st minimum value that up to the present objective function XB finds,

the 2nd minimum value that up to the present objective function Jm finds;

Step 7: multi-objective problem F (x)=min (F1 (x), F2 (x)) is resolved into N subproblem with Chebyshev's decomposition method, and the objective function of each concrete subproblem is as follows:

Wherein,

current reference point, the vector that the current optimal value of each target forms, in the present invention, the value of m is 2;

represent the objective function of j subproblem;

the weights of j subproblem;

x represents a population at individual, f _ji(x) value of i objective function of the individuality correspondence of j subproblem of expression;

Step 8: according to each subproblem weights λ ^j, calculate s_n neighbours subproblem Nbor (j)=(Nbor of each subproblem _j1, Nbor _j2..., Nbor _{js_n}), Nbor _jirepresent the index of i the neighbours subproblem of j subproblem, so Nbor _jivalue be integer; Get s_n=10; I=1,2 ..., s_n;

Step 9: by each subproblem

individual P _j(t) be initialized as x _j, x _j∈ X, wherein t is iterations, t=0; And calculate individual P _j(t) corresponding target function value Ft _j;

Step 10: to each subproblem

corresponding individual P _j(t) carry out evolutional operation and obtain interim individual P _j(t+1) "

3 neighbours subproblem s of random selection in 10.1 s_n j subproblem neighbours subproblem Nbor (j), k, l, to s, k, the individual P of l neighbours subproblem _s(t), P _k(t), P _l(t) simulate two and enter interlace operation, obtain a new interim offspring individual P _j(t+1) ';

10.2 couples of interim individual P _j(t+1) ' and carry out polynomial expression mutation operation, obtain individual P _j(t+1) ";

Step 11: to the individual P obtaining _j(t+1) " carry out an iterative operation with FCM algorithm and obtain new individual P _j(t+1);

Step 12: calculate new interim individual P _j(t+1) two target function value newF _j, and according to newF _jupgrade ideal point Z ^*; By new interim individual P _jand its desired value newF (t+1) _jupgrade all s_n neighbours subproblem Nbor (j) target function value of corresponding individual and each individual correspondence respectively of j subproblem;

Step 13: judge whether current iteration number of times t meets t<Tmax, as met, perform step 13; Otherwise, make iterations t add t=t+1 one time, return to step 11, wherein Tmax is maximum iteration time, gets Tmax=20;

Step 14: select suitable individuality as cluster centre from population: by each subproblem

the individual P of parent _j(t) take out, using each individual cluster centre as super pixel that will cluster, using these cluster centres as final output disaggregation; And concentrate and select one by one body as cluster centre from final output solution according to third party's index PBM;

Step 15: the classification number that obtains each pixel: the Euclidean distance that calculates the proper vector of each pixel and the cluster centre that obtains from step 15, this pixel is grouped in the classification of cluster centre of Euclidean distance minimum apart from it, obtains the classification of each pixel;

Step 16: image is cut apart in output.

2. a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM according to claim 1, is characterized in that: the process of the extraction characteristics of image to be split described in step 2 comprises as follows:

2.1 processes of utilizing Gabor wave filter to extract the medium and low frequency texture feature vector of image include: the Gabor kernel function of bidimensional can be defined as:

Wherein, σ _xand σ _yrepresent that respectively oval Gaussian function is along the standard deviation in x and y direction, F is modulating frequency, and θ is the direction of Gabor kernel function.

2.2 utilize the process of algorithm of co-matrix texture feature extraction vector to include: be first certain gray level by pending image quantization, local window size can design according to object of experiment, distance between pixel is d_n, the angular separation that makes successively again two pixel lines and transverse axis is certain angle, distinguishes according to the following formula the gray level co-occurrence matrixes of the direction number of calculation requirement:

P(i,j)=#{(x ₁,y ₁),(x ₂,y ₂)∈M×N|f(x ₁,y ₁)=r,f(x ₂,y ₂)=s}

Wherein, P (i, j) be gray level co-occurrence matrixes at the locational element of coordinate (i, j), # for set { } element number, (x ₁, y ₁) and (x ₂, y ₂) be two pixel coordinates apart from equaling d_n, ∈ is the symbol that belongs in set, the size that M × N is pending image, | be the conditional code in theory of probability, r is (x ₁, y ₁) locating the gray-scale value after pixel vector quantization, s is (x ₂, y ₂) locate the gray-scale value after pixel vector quantization.

3. a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM according to claim 1, is characterized in that: the computation process of calculating the target function value of each individuality described in step 5 includes:

The formula of the 5.1 target function value F1 that obtain according to XB index and Jm index respectively and F2 is as follows:

Wherein, features _jthe pixel characteristic matrix extracting, j=1,2 ..., Nf is the line number of extracting feature eigenmatrix afterwards, i.e. the number of the image block after image coarse segmentation, c _p, p=1,2 ... k is the cluster centre of image pixel, and k is the number of the cluster centre of appointment, U _{k × Nf}it is fuzzy membership matrix;

The target function value F of 5.2 each individualities _n=[F _n1, F _n2], wherein, F _n1=F1=XB, F _n2=F2=Jm.

4. a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM according to claim 1, is characterized in that: in step 11 to the individual P obtaining _j(t+1) " carry out an iterative operation with FCM algorithm and obtain new individual P _j(t+1) described implication is as follows:

11.1 individual P _j(t+1), " as cluster centre c, obtain the fuzzy membership matrix u of this cluster centre to the Euclidean distance of data features to be clustered according to this cluster centre c _ij, u _ijbe calculated as follows:

U _ijbetween 0,1, c _ithe cluster centre of dimmed group of i, d _ij=|| c _i-features _j|| represent the Euclidean distance of j the data point of i cluster centre and features, and m ∈ [1, ∞), m=2 in the present invention;

11.2 according to the 11.1 fuzzy membership matrix u that obtain _ijcluster centre c ' after must being upgraded, c ' is calculated as follows:

Features _jbe j data point, nf is the number of the data point that comprises of features;

11.3 c ' that obtain are as new individual P _j(t+1).

5. a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM according to claim 1, is characterized in that: step 12 ideal point Z ^*renewal process and more the process of the individual and corresponding desired value of the parent of new neighbor subproblem comprise as follows:

12.1 ideal point Z ^*renewal process include: if

otherwise constant; If

otherwise constant;

12.2 more all individualities of new neighbor subproblem and the process of corresponding desired value thereof include: for new interim individual P _j(t+1) each neighbours subproblem, Nbor _ji∈ Nbor (j), i=1,2 ..., s_n, wherein s_n is the number of neighbours subproblem, if for all Nbor _jihave

use new interim individual P _j(t+1) i the individuality that neighbours subproblem is corresponding of alternative j subproblem

and use newF _jsubstitute i the target function value that neighbours subproblem is corresponding of j subproblem

otherwise, constant.

6. a kind of SAR image partition method based on decomposing Evolutionary multiobjective optimization and FCM according to claim 1, is characterized in that: the process that the selection according to third party's index PBM described in step 14 produces optimum segmentation result is as follows:

First 14.1 obtain the value obtaining according to third party's index PBM of each cluster centre:

Wherein, features _j, c _p, u _pjconsistent with the description in step 2, index=1 ... N, N is the number of the current cluster centre obtaining, k is the class number that will cut apart, E _kthe super pixel that obtains of a splitting scheme and the distance sum D of its corresponding cluster centre _kbe the maximum separation between cluster centre, Ec is each features _j, j=1 ... Nf is to the distance sum of the geometric center of features, and each cluster centre obtains Ec and equates.

Images