CN104537660A - Image cutting method based on multi-target intelligent body evolution clustering algorithm - Google Patents

Abstract

The invention discloses an image cutting method based on a multi-target intelligent body evolution clustering algorithm. The problems that the image cutting technology is prone to local optimum and an algorithm is not high in robustness are mainly solved. The image cutting problem is converted into a global optimization clustering problem. The process includes the steps of extracting gray information of pixel points of an image to be cut, initiating parameters and establishing an image intelligent body network, calculating the energy of an image intelligent body, conducting non-domination sequencing, conducting neighborhood competition operation, conducting Gaussian mutation operation, calculating the energy of the image intelligent body, conducting non-domination sequencing, conducting self-learning operation, selecting the optimal clustering result according to the crowding distance, outputting a clustering label, and achieving image cutting. Multiple targeting is achieved for the image processing process, the convergence effect is good, the robustness of the method is enhanced, the image cutting quality can be improved, the cutting effect stability can be enhanced, and the extraction, recognition and other subsequent processing of the image targets are facilitated.

Description

Based on the image partition method of Multiobjective Intelligent body evolution clustering algorithm

Technical field

The present invention is under the jurisdiction of technical field of image processing, relates generally to image partition method, and specifically a kind of image partition method based on Multiobjective Intelligent body evolution clustering algorithm, can be used for the field such as pattern-recognition and computer vision.

Background technology

Image is a gordian technique in image processing, it has a very wide range of applications in image procossing research, such as to target identification, target measurement all based on Iamge Segmentation, segmentation result directly has influence on the carrying out of follow-up work, and therefore the research tool of Iamge Segmentation is of great significance.Iamge Segmentation is a kind of special image processing techniques in fact, and it is in fact carry out the process of classifying according to image pixel attribute and gray scale, texture, color.

Method based on cluster is the one of unsupervised segmentation, is widely used in the fields such as work biomedicine, computer vision and remote sensing image processing.Cluster is classified at one of unknown distribution group of data, farthest make the data in same classification have identical character, and inhomogeneous data has different character, and it is in fact hide rule in the data to find.

Based on this, various clustering algorithm is applied to Iamge Segmentation field and also achieves more and more satisfied effect.But due to singularity and the diversity of view data, and the clustering algorithm of not all can be applied directly to Iamge Segmentation field, a lot of algorithm all needs to carry out packaging to be improved, and a part of algorithm is not suitable for carrying out Iamge Segmentation even.In current research, conventional clustering technique has following several: hierarchical clustering algorithm, nearest-neighbor clustering algorithm, fuzzy clustering algorithm, artificial neural network clustering algorithm, genetic algorithm for clustering.Initial research personnel are usually hierarchical clusters for the cluster of Iamge Segmentation.Its advantage is simple, is easy to operation, but simultaneously it also brings a lot of inconvenience: as, Algorithm robustness is poor, is easily absorbed in local optimum, and cluster result is not very desirable etc.In order to solve this kind of problem, many researchists have carried out a lot of trial, propose to adopt Genetic Algorithms combined with it, obtain satisfied result, but due to the limitation of traditional genetic algorithm overall situation evolution mechanism, in conjunction with after clustering method still to have cluster result unstable and be easily absorbed in the defects such as local optimum, cause the Quality Down of image segmentation result and the reduction of segmentation effect stability, be unfavorable for follow-up graphical analysis and understanding.

Summary of the invention

The object of the invention is to overcome and be above-mentionedly easily absorbed in local optimum and the not high deficiency of Algorithm robustness, a kind of image partition method based on Multiobjective Intelligent body evolution clustering algorithm is proposed, with the stability of the robustness and image segmentation result that strengthen dividing method, improve image segmentation.

Technical scheme of the present invention is, based on the image partition method of Multiobjective Intelligent body evolution clustering algorithm, it is characterized in that: include following steps:

Step 1 inputs image to be split, extracts the half-tone information of image to be split, and need carry out pre-service if input picture is not gray level image it is changed into gray level image, half-tone information is labeled as data.

Step 2 sets initial clustering number c and iteration upper limit number of times T, gradation of image information is carried out to the random initializtion of clustering prototype, clustering prototype is the initial clustering result of gradation of image information, the random initializtion of clustering prototype refers to that random appointment c pixel value is cluster centre, by the pixel classifications of image to margin of image element apart from minimum cluster centre, the clustering prototype of each half-tone information as an image intelligent body, then for image intelligent body L _y, the subscript y of L refers to the numbering of image intelligent body, y=1,2 ..., L _size× L _size, L _y={ e _p, e _pimage intelligent body L _yin element, p is element subscript in image intelligent body, 1≤p≤c, setting image intelligent body existence grid L ^tsize is L _size× L _size, competition probability is P ₀, mutation probability is P _m, make current evolutionary generation t=0.

Step 3 computed image intelligent body L _yeNERGY E rengyL _y=(E ₁, E ₂), y refers to the numbering of image intelligent body, y=1,2 ..., L _size× L _size, ENERGY E rengyL _ycomprise energy 1 component L _y(E ₁) and energy 2 component L _y(E ₂).

Step 4 according to the energy value of all image intelligent bodies, to image intelligent volume mesh L ^tmiddle image intelligent body L _ycarry out non-dominated ranking operation, the image intelligent body that order is not arranged by other image intelligent body any is ground floor, is saved in set level1, makes all the other all image intelligent bodies be saved in set level2;

Neighborhood is competed on the image intelligent body that operator acts in set level2 by step 5, and the new image intelligent body obtained and the image intelligent body in set level1 is merged and jointly form new image intelligent volume mesh L ^t+1/2.

Step 6 initiatively produces the random number R (0,1) of 0 to 1, by mutation probability P _mcompare, if mutation probability p with R (0,1) _mbe greater than R (0,1), then Gaussian mutation operator acted on successively image intelligent volume mesh L ^t+1/2in intelligent body on, obtain image intelligent volume mesh L of future generation ^t+1.

Step 7 is to image intelligent volume mesh L of future generation ^t+1middle image intelligent body performs step 3 to step 4, obtains one group of non-dominant image intelligent body L _a, a refers to the numbering of non-dominant image intelligent body, a=1,2 ..., A, 1≤A≤L _size× L _size, this group non-dominant image intelligent body refers to the image intelligent body in ruleless ground floor.

Step 8 couple non-dominant image intelligent body L _acarry out self-learning operator operation successively, obtain second generation non-dominant disaggregation, comprise second generation non-dominant image intelligent body L _b, b refers to the numbering of second generation non-dominant image intelligent body, b=1,2 ..., A, 1≤A≤L _size× L _size;

Step 9 judges whether iteration terminates: if evolutionary generation t is not less than stop algebraically T, jump procedure 10, if evolutionary generation t is less than stop algebraically T, t=t+1, returns step 3;

Step 10 is according to the second generation non-dominant image intelligent body L obtained in step 8 _bb=1,2 ..., A, 1≤A≤L _size× L _size, calculate the crowding distance dis of each image intelligent body successively,

$\mathrm{dis}\left(L_b\right)={(L_{b-1}\left(E_1\right)-L_{b+1}\left(E_1\right))}^2+{(L_{b-1}\left(E_2\right)-L_{b+1}\left(E_2\right))}^2$

In formula, b=1,2,3 ..., A, selects optimum image intelligent body L to crowding distance according to maximum crowding distance principle _bestand export optimum cluster label corresponding to this image intelligent body;

Step 11 carries out pixel classifications according to optimum cluster label to input picture, and using similar pixel as a block of pixels, all block of pixels press image original position composition segmentation image, export the segmentation result of input picture.

Compared with prior art, advantage of the present invention is the restriction overcoming traditional images segmentation, achieves:

1) because the present invention has used for reference the algorithm frame of multiple goal thought, in cluster process, design employs the operations such as quick non-dominated ranking operator and crowding distance selection, overcome optimal result in iterative process single, the shortcomings such as population sample is single, make the population of image segmentation process more diversified, segmentation result is closer to optimal result.

2) because present invention employs the operation such as Agent Grid structure, neighborhood competition operator, Gaussian mutation operator and self-learning operator, therefore overcome and be easily absorbed in the shortcoming such as local optimum and segmentation result instability, make the accuracy of image segmentation result higher, also improve the stability of image segmentation.

The population diversification more of cluster process in Iamge Segmentation is can be implemented in by the present invention, and the limitation of normal image dividing method has been broken in this invention, initiative combines multiple goal thought with multi-Agent evolutionary Algorithm, propose a kind of evolution clustering algorithm based on Multiobjective Intelligent body, achieve the raising to segmentation result accuracy, and enhance the stability of segmentation effect.

Accompanying drawing explanation

The realization flow figure of Fig. 1 the inventive method;

The grid environment of Fig. 2 intelligent body existence;

Fig. 3 the inventive method is applied to the emulation segmentation result of two gray-like image pictures;

Fig. 4 the inventive method, Kmeans cluster segmentation and FCM cluster segmentation are applied to the simulation result of sar river image;

Fig. 5 the inventive method, Kmeans cluster segmentation and FCM cluster segmentation are applied to the simulation result of sar Airport Images;

Embodiment

Below in conjunction with accompanying drawing to the detailed description of the invention:

Based on the image partition method of Multiobjective Intelligent body evolution clustering algorithm, its committed step is the combination of multiple goal thought and multi-Agent evolutionary Algorithm, and completes Iamge Segmentation for cluster optimization.

Feature based on the image partition method of Multiobjective Intelligent body evolution clustering algorithm is: the half-tone information first extracting image to be split, then with Multiobjective Intelligent body evolution algorithm cluster carried out to it and optimize, then according to the final cluster labels exported pixel classified and then realize the segmentation of image.

Embodiment 1

With reference to Fig. 1, the present invention is a kind of image partition method based on Multiobjective Intelligent body evolution clustering algorithm, and specific implementation step is as follows:

Step 1, inputs image to be split, extracts the half-tone information of image to be split, and need carry out pre-service if input picture is not gray level image it is changed into gray level image, half-tone information is labeled as data.

Step 2, setting initial clustering number c and iteration upper limit number of times T, gradation of image information is carried out to the random initializtion of clustering prototype, clustering prototype is the initial clustering result of gradation of image information, the random initializtion of clustering prototype refers to specifies c pixel value to be cluster centre at random, is saved in image intelligent body, by the pixel classifications of image to margin of image element apart from minimum cluster centre, each clustering prototype as an image intelligent body, then for image intelligent body L _y, the subscript y of L refers to the numbering of image intelligent body, y=1,2 ..., L _size× L _size, L _y={ e _p, e _pimage intelligent body L _yin element, p is element subscript in image intelligent body, 1≤p≤c, setting image intelligent body existence grid L ^t, t refers to current evolutionary generation, and sizing grid is L _size× L _size, competition probability is P ₀, mutation probability is P _m, make current evolutionary generation t=0; In this example, clusters number c=2, iteration upper limit number of times T=100, Agent Grid size L _size=8, competition probability P ₀=0.2, Gaussian mutation probability P _m=0.1;

Image intelligent volume mesh is a grid environment of image intelligent body existence, and be designated as L, size definition is L _size× L _size, wherein L _sizefor integer, each image intelligent body is fixed on a lattice point, and it is L that note is in the image intelligent body that the i-th i is capable, jth j arranges _{ii, jj}, ii, jj=1,2 ..., L _size, image intelligent body L _{ii, jj}neighborhood be:

$L_{\mathrm{ii},\mathrm{jj}}^n=\{L_{{\mathrm{ii}}^{\′},\mathrm{jj}},L_{\mathrm{ii},{\mathrm{jj}}^{\′}},L_{{\mathrm{ii}}^{\′\′},\mathrm{jj}},L_{\mathrm{ii},{\mathrm{jj}}^{\′\′}}\}$

Wherein,

${\mathrm{ii}}^{\′}=\left\{\begin{array}{cc}\mathrm{ii}-1& \mathrm{ii}\≠1\\ L_{\mathrm{size}}& \mathrm{ii}=1\end{array}\right.,{\mathrm{jj}}^{\′}=\left\{\begin{array}{cc}\mathrm{jj}-1& \mathrm{jj}\≠1\\ L_{\mathrm{size}}& \mathrm{jj}=1\end{array}\right.,{\mathrm{ii}}^{\′\′}=\left\{\begin{array}{cc}\mathrm{ii}+1& \mathrm{ii}\≠L_{\mathrm{size}}\\ 1& \mathrm{ii}=L_{\mathrm{size}}\end{array}\right.,{\mathrm{jj}}^{\′\′}=\left\{\begin{array}{cc}\mathrm{jj}+1& \mathrm{jj}\≠L_{\mathrm{size}}\\ 1& \mathrm{jj}=L_{\mathrm{size}}\end{array}\right.$

Each image intelligent body can not move, and can only interact with its neighborhood, for image intelligent body L _{ii, jj}the image intelligent body that neighboring region energy is maximum;

Fig. 2 gives the structural representation of image intelligent volume mesh, and in figure, each circle represents an image intelligent body, this image intelligent body position within a grid of the numeral in circle, and has two of mutual line image intelligent bodies to interact.

Step 3, utilizes all image intelligent body L in following two formulae discovery populations _yeNERGY E rengyL _y=(E ₁, E ₂), the subscript y of L refers to the numbering of image intelligent body, y=1,2 ..., L _size× L _size, computation process comprises 1 component L _y(E ₁) calculating and 2 component L _y(E ₂) calculating, wherein: image intelligent body L _y1 component of energy calculates and is

$L_y\left(E_1\right)-\frac{1}{\underset{j=1}{\overset{c}{\mathrm{\Σ}}}\underset{x\∈S_j}{\mathrm{\Σ}}{\left|\right|x-z_j\left|\right|}^2}$

In formula, x is cluster image intensity value, z _jfor corresponding cluster centre, image intelligent body L _y2 components of energy calculate and are

In formula, i is a data point, and m is the number of data point, and j is the nearest neighbor point of i, and s is nearest neighbor point number, x _i,jbe the relation value of i-th data point and a jth nearest neighbor point, wherein i, j are similar pixel then value 0, and inhomogeneity is then got

Step 4, according to the energy value of all image intelligent bodies, to image intelligent volume mesh L ^tmiddle image intelligent body, y=1,2 ..., L _size× L _sizecarry out non-dominated ranking operation, order is not by the image L of any other image intelligent body domination _yintelligent body is ground floor, is saved in set level1, makes all the other all image intelligent bodies be saved in set level2.Detailed process comprises:

(4.1) initiation parameter, order domination intelligent body L _yintelligent body number n _ly=0.

(4.2) n is upgraded by following formula _ly:

For image intelligent body L _y, y=1,2 ..., L _size× L _size, have image intelligent body, i refers to the numbering of image intelligent body, i=1,2 ..., L _size× L _size, and L _ican not be L _yif, image intelligent L _il _yenergy meet $\left\{\begin{array}{c}{L}_y\left(E_1\right)<L_i\left(E_1\right)\\ L_y\left(E_2\right)<L_i\left(E_2\right)\end{array}\right.,$ Then image intelligent body L _yby image intelligent body L _iarranged, made n _ly=n _ly+ 1, if meet $\left\{\begin{array}{c}{L}_y\left(E_1\right)<L_i\left(E_1\right)\\ L_y\left(E_2\right)<L_i\left(E_2\right)\end{array}\right.,$ Then claim image intelligent body L _ydominating figure is as intelligent body L _i.

(4.3) if image intelligent body L _y, y=1,2 ..., L _size× L _sizen _ly=0 makes image intelligent body L _yenter ground floor, be saved in set level1, have n _level1individuality, if its n _ly≠ 0, make L _yenter the second layer, be saved in level2.The image intelligent body of different levels is separately preserved by the present invention, is conveniently follow-up use.

Step 5, competes neighborhood on image intelligent body that operator acts in set level2 successively, the new image intelligent body obtained and the image intelligent body in set level1 is merged, the image intelligent volume mesh L that after merging, common composition is new ^t+1/2.

Step 6, initiatively produces the random number R (0,1) of 0 to 1, compares, mutation probability and random number R (0,1) if mutation probability p _mbe greater than random number R (0,1), then Gaussian mutation operator acted on successively image intelligent volume mesh L ^t+1/2in intelligent body on, obtain image intelligent volume mesh L of future generation ^t+1.

Gaussian mutation operator produces a variation image intelligent body by following formula

Wherein p=1,2 ..., c, c are clusters number, l _pfor intelligent body L _{ii, jj}in be positioned at the element value at p place, G (0,1/t) is the random number of Gaussian distribution; And random number R (0,1) is the random number between 0 to 1, T is total evolutionary generation, and t is current evolutionary generation.

Step 7, to image intelligent volume mesh L of future generation ^t+1middle image intelligent body performs step 3 to step 4, obtains one group of non-dominant image intelligent body L _a, a refers to the numbering of non-dominant image intelligent body, a=1,2 ..., A, 1≤A≤L _size× L _size, this group non-dominant image intelligent body refers to the image intelligent body in ruleless ground floor.

Step 8, to non-dominant image intelligent body L _a, a=1,2 ... A, 1≤A≤L _size× L _sizecarry out self-learning operator operation successively, obtain second generation non-dominant disaggregation, comprise second generation non-dominant image intelligent body L _b, b refers to the numbering of second generation non-dominant image intelligent body, b=1,2 ..., A, 1≤A≤L _size× L _size.

Step 9, judges whether iteration terminates: if evolutionary generation t is not less than stop algebraically T, and jump procedure (10), if evolutionary generation t is less than stop algebraically T, t=t+1, returns step 3.

Step 10, according to the second generation non-dominant image intelligent body L obtained in step 8 _b, calculate the crowding distance dis of each image intelligent body successively,

$\mathrm{dis}\left(L_b\right)={(L_{b-1}\left(E_1\right)-L_{b+1}\left(E_1\right))}^2+{(L_{b-1}\left(E_2\right)-L_{b+1}\left(E_2\right))}^2$

In formula, b=1,2,3 ..., A, selects optimum image intelligent body L to crowding distance according to maximum crowding distance principle _bestand export optimum cluster label corresponding to this image intelligent body.

Step 11, carries out pixel classifications according to optimum cluster label to input picture, and using similar pixel as a block of pixels, all block of pixels press image original position composition segmentation image, export the segmentation result of input picture.

Because the present invention has used for reference the algorithm frame of multiple goal thought, in cluster process, design employs the operations such as quick non-dominated ranking operator and crowding distance selection, overcome optimal result in iterative process single, the shortcomings such as population sample is single, make the population of image segmentation process more diversified, segmentation result is closer to optimal result.

Embodiment 2

Based on the image partition method of Multiobjective Intelligent body evolution clustering algorithm with embodiment 1, wherein in step 5, form new image intelligent volume mesh L ^t+1/2process be neighborhood is competed on image intelligent body that operator acts in set level2 successively, and the new image intelligent body obtained and the image intelligent body in set level1 to be merged, after merging, jointly form new image intelligent volume mesh L ^t+1/2.Neighborhood competition process produces new image intelligent body by neighborhood competition operator with two kinds of strategies, and total process comprises:

Neighborhood competition operator is by the new image intelligent body of one of the following two kinds strategy generation one

Strategy 1, produces image intelligent body by following formula

Wherein p=1,2 ..., c, e _pfor in element, for the lower bound of all image intelligent element of volume values, for the upper bound of all image intelligent element of volume values, m _pfor image intelligent body arbitrary in level1 is positioned at the element value at p place, l _pfor image intelligent body L _{ii, jj}in be positioned at the value at p place, c is clusters number, and R (-1,1) is the random number between-1 to+1.

Strategy 2, produces image intelligent body as follows

The first step, will by following formula middle all elements m _pbe mapped on interval [0,1], obtain new unit

Element: for image intelligent body in the middle of forming with these elements

$L_{\mathrm{ii},\mathrm{jj}}^{{\mathrm{new}}^{\&prime;}}=(m_1^{\&prime;},m_2^{\&prime;},...,m_{i_1-1}^{\&prime;},m_{i_2-1}^{\&prime;},...,m_{i_1+1}^{\&prime;},m_{i_1}^{\&prime;},m_{i_2+1}^{\&prime;},m_{i_2+2}^{\&prime;},...,m_p^{\&prime;})$

Wherein p=1,2 ..., c, 1<i ₁<c, 1<i ₂<c, i ₁<i ₂, c is optimum cluster number, m _pfor image intelligent body arbitrary in level1 is positioned at the element value at p place, for the lower bound of all image intelligent element of volume values, for the upper bound of all image intelligent element of volume values;

Second step, according to following formula by image intelligent body map go back to interval on, obtain image intelligent body $L_{\mathrm{ii},\mathrm{jj}}^{\mathrm{new}}=\left\{e_p\right\}:$

$e_p={\underset{\&OverBar;}{x}}_p+m_p^{\&prime;}({\stackrel{\&OverBar;}{x}}_p-{\underset{\&OverBar;}{x}}_p),p=\mathrm{1,2},...c$

Wherein, e _pfor in element, for the lower bound of all image intelligent element of volume values, for the upper bound of all image intelligent element of volume values, c is clusters number.

Two kinds of above-mentioned Different Strategies are according to competition probability P ₀select:

First, the random number R (0,1) between 0 to 1 is initiatively produced;

Secondly, random number R (0,1) and competition probability P is made ₀compare, if random number R (0,1) > P ₀, then selection strategy 1, otherwise, selection strategy 2.

Embodiment 3

Based on the image partition method of Multiobjective Intelligent body evolution clustering algorithm with embodiment 1-2, wherein in step 8, obtain second generation non-dominant image intelligent body L _b, b is the numbering of second generation non-dominant image intelligent body, b=1,2 ..., A, 1≤A≤L _size× L _sizeprocess be by non-dominant image intelligent body L _a, a=1,2 ..., A, 1≤A≤L _size× L _sizecarry out self-learning operator operation successively to obtain.Self-learning operator operating process produces a new image intelligent body by self-learning operator, and total process comprises:

Self-learning operator produces an image intelligent body as follows:

(8.1) method utilizing image intelligent volume mesh to generate produces a self study image intelligent volume mesh sL, and its size is sL _size× sL _size, sL _sizefor integer, all image intelligent body sL on it _{i ', j '}, i ', j '=1,2 ..., sL _sizeproduce according to following formula:

Wherein p=1,2 ..., c, for the lower bound of all image intelligent body gray-scale values; for the upper bound of all image intelligent body gray-scale values; l _pfor intelligent body L _a, a=1,2 ... A, 1≤A≤L _size× L _sizein be positioned at the gray-scale value at p place, R (1-sR, 1+sR) represents the random number between 1-sR to 1+sR, and sR ∈ [0,1] represents search radius.

(8.2) neighborhood is competed operator and mutation operator iteration acts on self study image intelligent volume mesh sL, greatest iteration algebraically is sG, and the image intelligent body maximum with energy in self study image intelligent volume mesh sL substitutes present image intelligent body L _a, a=1,2 ... A, 1≤A≤L _size× L _size.

In this example, sL _size=4, sR=0.3, sG=8.

Because present invention employs the operation such as Agent Grid structure, neighborhood competition operator, Gaussian mutation operator and self-learning operator, therefore overcome and be easily absorbed in the shortcoming such as local optimum and segmentation result instability, make the accuracy of image segmentation result higher, also improve the stability of image segmentation.

Effect of the present invention can be further illustrated by following simulation result.

Embodiment 4

Based on the image partition method of Multiobjective Intelligent body evolution clustering algorithm with embodiment 1-3, by emulation experiment, implementation procedure of the present invention and technique effect are further illustrated.

Optimum configurations:

Clusters number c=2, maximum evolutionary generation T=100, Agent Grid size L _size=8, competition probability P ₀=0.2, Gaussian mutation probability P _m=0.1, self study Agent Grid size sL _size=4, search radius sR=0.3, self study mutation probability sP _m=0.1, self study algebraically sG=8.

Emulation content:

Apply image partition method of the present invention, respectively to a width two gray-like image picture and a width SAR river Image Segmentation Using, and give result to illustrate that the inventive method is applied to the performance of Iamge Segmentation.

Emulation content 1, image partition method of the present invention is applied to two gray-like image pictures and splits, its result is as shown in Figure 8, wherein: Fig. 3 (a) is original composograph, object gray-scale value is 255, and background gray levels is 51, Fig. 3 (b) is segmentation result figure.Can find out from result figure and present invention obtains correct segmentation result, successfully be partitioned into the circular object in two gray-like image pictures of synthesis, illustrated that the inventive method can obtain correct segmentation result.

Emulation content 2, image partition method of the present invention is applied to the emulation experiment of SAR river Image Segmentation Using, and the emulation experiment of conventional segmentation methods Kmeans cluster segmentation method and FCM cluster segmentation method, its result as shown in Figure 4, Fig. 4 (a) is former SAR river image, Fig. 4 (b) is the segmentation result figure of Kmeans cluster segmentation method, the segmentation result figure that Fig. 4 (c) is dividing method of the present invention for the segmentation result figure of FCM cluster segmentation method, Fig. 4 (d).As can be seen from Fig. 4 (d), for SAR river image, the segmentation result of this method has good region consistency, compare traditional cluster segmentation method Kmeans cluster segmentation result and the middle FCM cluster segmentation result of Fig. 4 (d) in Fig. 4 (c), the inventive method successfully can not only be partitioned into the river in image, the mistake of land buildings is divided and also reduces a lot, greatly reduce the impact that noise spot is split river.

Embodiment 5

Based on the image partition method of Multiobjective Intelligent body evolution clustering algorithm with embodiment 1-3, simulated conditions and simulation parameter arrange same embodiment 4.

Emulation content 3, image partition method of the present invention is applied to the emulation experiment that SAR Airport Images carries out splitting, and the emulation experiment of conventional segmentation methods Kmeans cluster segmentation method and FCM cluster segmentation method, its result as shown in Figure 5, Fig. 5 (a) is former SAR river image, Fig. 5 (b) is the segmentation result figure of Kmeans cluster segmentation method, the segmentation result figure that Fig. 5 (c) is dividing method of the present invention for the segmentation result figure of FCM cluster segmentation method, Fig. 5 (d).Find out from Fig. 5 (d) simulation result, for SAR airfield runway image, compare the Kmeans cluster segmentation result in Fig. 5 (b) and the FCM cluster segmentation result in Fig. 5 (c), the segmentation result of the inventive method has better region consistency, and the impact that inhibit the noise on image on airfield runway to split,, the weak signal target except primary runway is then ignored meanwhile, be more clearly partitioned into airport primary runway.

In brief, the image partition method based on Multiobjective Intelligent body evolution clustering algorithm of the present invention, mainly solves in conventional images cutting techniques and is easily absorbed in the problems such as local optimum, Algorithm robustness are not high.The segmentation problem of image is converted into the clustering problem of a global optimization by the present invention, and its segmentation step is: the half-tone information extracting image slices vegetarian refreshments to be split; Parameter initialization also sets up image intelligent volume mesh; Calculate the energy of all image intelligent bodies and carry out non-dominated ranking; Successively neighborhood competition operator operation is carried out to second layer image intelligent body and and ground floor image intelligent body merge and form new image intelligent volume mesh; Carry out Gaussian mutation operation to new Agent Grid to calculate; Again non-dominated ranking is carried out to the energy of all image intelligent bodies; Carry out self study operation; Select optimum cluster result according to crowding distance, export cluster labels; Realize Iamge Segmentation.The present invention is by the processing procedure multiple goal of Iamge Segmentation, not only good in convergence effect, and enhance the robustness of method, so the quality of Iamge Segmentation can be improved and strengthen the stability of segmentation effect, be conducive to the extraction of image object, identification and some other subsequent treatment.

Claims (6)

1. based on an image partition method for Multiobjective Intelligent body evolution clustering algorithm, it is characterized in that, include following steps:

Step 1 inputs image to be split, and extract the half-tone information of image to be split, half-tone information is labeled as data;

Step 2 sets initial clustering number c and iteration upper limit number of times T, gradation of image information is carried out to the random initializtion of clustering prototype, the clustering prototype of each half-tone information as an image intelligent body, then for image intelligent body L _y, the subscript y of L refers to the numbering of image intelligent body, y=1,2 ..., L _size× L _size, L _y={ e _p, e _pimage intelligent body L _yin element, p is element subscript in image intelligent body, 1≤p≤c, setting image intelligent body existence grid L ^tsize is L _size× L _size, competition probability is P ₀, mutation probability is P _m, make current evolutionary generation t=0;

Step 3 computed image intelligent body L _yeNERGY E rengyL _y=(E ₁, E ₂), the subscript y of L refers to the numbering of image intelligent body, y=1,2 ..., L _size× L _size, ENERGY E rengyL _ycomprise energy 1 component L _y(E ₁) and energy 2 component L _y(E ₂);

Step 4 according to the energy value of all image intelligent bodies, to image intelligent volume mesh L ^tmiddle image intelligent body L _ycarry out non-dominated ranking operation, the image intelligent body that order is not arranged by other image intelligent body any is ground floor, is saved in set level1, makes all the other all image intelligent bodies be saved in set level2;

Neighborhood is competed on the image intelligent body that operator acts in set level2 successively by step 5, and the new image intelligent body obtained and the image intelligent body in set level1 is merged and jointly form new image intelligent volume mesh L ^t+1/2;

Step 6 initiatively produces the random number R (0,1) of 0 to 1, by mutation probability P _mcompare, if mutation probability p with R (0,1) _mbe greater than R (0,1), then Gaussian mutation operator acted on successively image intelligent volume mesh L ^t+1/2in intelligent body on, obtain image intelligent volume mesh L of future generation ^t+1;

Step 7 is to image intelligent volume mesh L of future generation ^t+1middle image intelligent body performs step 3 to step 4, obtains one group of non-dominant image intelligent body L _a, a refers to the numbering of non-dominant image intelligent body, a=1,2 ..., A, 1≤A≤L _size× L _size;

Step 8 couple non-dominant image intelligent body L _acarry out self-learning operator operation successively, obtain second generation non-dominant disaggregation, comprise second generation non-dominant image intelligent body L _b, b refers to the numbering of second generation non-dominant image intelligent body, b=1,2 ..., A, 1≤A≤L _size× L _size;

Step 9 judges whether iteration terminates: if evolutionary generation t is not less than stop algebraically T, jump procedure 10, if evolutionary generation t is less than stop algebraically T, t=t+1, returns step 3;

Step 10 is according to the second generation non-dominant image intelligent body L obtained in step 8 _b, calculate the crowding distance dis of each image intelligent body successively,

$\mathrm{dis}\left(L_b\right)=\sqrt{{(L_{b-1}\left(E_1\right)-L_{b+1}\left(E_1\right))}^2+{(L_{b-1}\left(E_2\right)-L_{b+1}\left(E_2\right))}^2}$

In formula, b=1,2,3 ..., A, selects optimum image intelligent body L to crowding distance according to maximum crowding distance principle _bestand export optimum cluster label corresponding to this image intelligent body;

Step 11 carries out pixel classifications according to optimum cluster label to input picture, and using similar pixel as a block of pixels, all block of pixels composition segmentation images, export the segmentation result of input picture.

2. the image partition method based on Multiobjective Intelligent body evolution clustering algorithm according to claim 1, is characterized in that, the calculating of image intelligent physical efficiency in step 3, carries out as follows:

(3a) computed image intelligent body L _yeNERGY E rengyL _y=(E ₁, E ₂) in energy 1 component L _y(E ₁),

$L_y\left(E_1\right)=\frac{1}{\underset{j=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{x\&Element;S_j}{\mathrm{\&Sigma;}}{\left|\right|x-z_j\left|\right|}^2}$

In formula, x is the half-tone information of gray level image, S _jfor corresponding cluster set, z _jfor the cluster centre of corresponding cluster set;

(3b) computed image intelligent body L _yeNERGY E rengyL _y=(E ₁, E ₂) in energy 2 component L _y(E ₂),

In formula, i is a pixel, and m is the number of pixel, and j is the nearest neighbor pixels point of i, and s is nearest neighbor pixels point number, x _i,jbe the relation value of a jth nearest neighbor pixels point of i-th pixel, wherein the then value 0 when pixel i, j belong to same cluster set, does not belong to same cluster set and then gets 1/j.

3. the image partition method based on Multiobjective Intelligent body evolution clustering algorithm according to claim 1, is characterized in that, the non-dominated ranking operation in step 4, carries out as follows:

(4a) initiation parameter, order domination intelligent body L _yintelligent body number n _ly=0;

(4b) n is upgraded by following formula _ly:

For image intelligent body L _y, y=1,2 ..., L _size× L _size, have image intelligent body L _i, i refers to the numbering of image intelligent body, i=1,2 ..., L _size× L _size, and L _ican not be L _yif, image intelligent L _yenergy meet $\left\{\begin{array}{c}{L}_y\left(E_1\right)<L_i\left(E_1\right)\\ L_y\left(E_2\right)<L_i\left(E_2\right)\end{array}\right.,$ Then image intelligent body L _yby image intelligent body L _iarranged, made n _ly=n _ly+ 1, if meet $\left\{\begin{array}{c}{L}_y\left(E_1\right)>L_i\left(E_1\right)\\ L_y\left(E_2\right)>L_i\left(E_2\right)\end{array}\right.,$ Then claim image intelligent body L _ydominating figure is as intelligent body L _i;

If (4c) image intelligent body L _y, y=1,2 ..., L _size× L _sizen _ly=0 makes image intelligent body L _yenter ground floor, be saved in set level1, have n _level1individuality, if its n _ly≠ 0, make L _yenter the second layer, be saved in level2.

4. the image partition method based on Multiobjective Intelligent body evolution clustering algorithm according to claim 1, is characterized in that, the neighborhood competition operator in step 5 produces an image intelligent body by one of the following two kinds strategy

Strategy 1: produce new image intelligent body by following formula

Wherein p=1,2 ..., c, e _pfor in element value, x _pfor the lower bound of element value in all image intelligent bodies, for the upper bound of element value in all image intelligent bodies, m _pfor image intelligent body in be positioned at the element value at p place, l _pfor image intelligent body L _{ii, jj}in be positioned at the element value at p place, c is clusters number, and R (-1,1) is the random number between-1 to+1;

Strategy 2, produces new image intelligent body as follows

The first step, will by following formula middle all elements value m _pbe mapped on interval [0,1], obtain new element: register map is formed as intelligent body with these elements

$L_{\mathrm{ii},\mathrm{jj}}^{{\mathrm{new}}^{\&prime;}}=(m_1^{\&prime;},m_2^{\&prime;},...,m_{i_1-1}^{\&prime;},m_{i_2-1}^{\&prime;},...,m_{i_1+1}^{\&prime;},m_{i_1}^{\&prime;},m_{i_2+1}^{\&prime;},m_{i_2+2}^{\&prime;},...,m_p^{\&prime;}),$

Wherein p=1,2 ..., c, 1<i ₁<c, 1<i ₂<c, i ₁<i ₂, c is clusters number, m _pfor image intelligent body in be positioned at the element value at p place, x _pfor the lower bound of all image intelligent element of volume values, for the upper bound of all image intelligent element of volume values;

Second step, according to following formula by image intelligent body map go back to interval on, obtain new image intelligent body $L_{\mathrm{ii},\mathrm{jj}}^{\mathrm{new}}=\left\{e_p\right\}:$

$e_p={\underset{\&OverBar;}{x}}_p+m_p^{\&prime;}({\stackrel{\&OverBar;}{x}}_p-{\underset{\&OverBar;}{x}}_p),p=\mathrm{1,2},...,c$

E _pfor in element value, x _pfor the lower bound of element value in all image intelligent bodies, for the upper bound of element value in all image intelligent bodies, c is clusters number;

Two kinds of described Different Strategies are according to competition probability P ₀select:

Random number R (0,1) by 0 to 1 and competition probability P ₀compare, if R (0,1) >P ₀, then selection strategy 1, otherwise, selection strategy 2.

5. the image partition method based on Multiobjective Intelligent body evolution clustering algorithm according to claim 1, is characterized in that, the Gaussian mutation operator in step 6, refers to produce a new variation image intelligent body by following formula $L_{\mathrm{ii},\mathrm{jj}}^{\mathrm{mut}}=\left\{e_p^{\mathrm{mut}}\right\},$

Wherein p=1,2 ..., c, c are clusters number, l _pfor intelligent body L _{ii, jj}in be positioned at the element value at p place; G (0,1/t) is the random number of Gaussian distribution; R (0,1) is the random number between 0 to 1, and T is total evolutionary generation, and t is current evolutionary generation.

6. the image partition method based on Multiobjective Intelligent body evolution clustering algorithm according to claim 1, is characterized in that, the self-learning operator in step 8 refers to the image intelligent body that generation one is new as follows:

(8a) method utilizing image intelligent volume mesh to generate produces a self study image intelligent volume mesh sL, and its size is sL _size× sL _size, sL _sizefor integer, all image intelligent body sL on it _{i ', j '}, i ', j '=1,2 ..., sL _sizeproduce according to following formula:

Wherein p=1,2 ..., c, x _pfor the lower bound of element value in all image intelligent bodies; for the upper bound of element value in all image intelligent bodies; l _pfor intelligent body L _a, a=1,2 ... A, 1≤A≤L _size× L _sizein be positioned at the gray-scale value at p place, R (1-sR, 1+sR) represents the random number between 1-sR to 1+sR, and sR ∈ [0,1] represents search radius;

(8b) neighborhood is competed operator and mutation operator iteration acts on self study image intelligent volume mesh sL, greatest iteration algebraically is sG, and the image intelligent body maximum with energy in self study image intelligent volume mesh sL substitutes present image intelligent body L _a, a=1,2 ... A, 1≤A≤L _size× L _size.