CN101866489A - Image dividing method based on immune multi-object clustering - Google Patents

Abstract

The invention discloses an image dividing method based on immune multi-object clustering, relating to the technical field of image processing, and mainly solving the problems that the conventional method has single evaluation index, and easily has bad region consistency and disorder boundary. The method comprises the following realization steps of: (1) extracting the characteristic of an image to be divided, and primarily dividing the image by controlling the watershed of a mark; (2) setting a running parameter and initializing the population of an antibody; (3) combining the locally-searched immune multi-objective optimizing method with the population of the antibody to obtain an approximate Pareto solution set; (4) selecting an optimal solution in the approximate Pareto solution set obtained in the step (3) according to a PBM index; and (5) marking an image pixel point according to a primary dividing result obtained in the step (1) and a clustering result obtained in the step (4) to obtain a final classifying result. The image dividing method has the advantages of good dividing result region consistency, being capable of keeping complete information, and having fast computation speed, and can be used for identifying an image object.

Description

Image partition method based on immune multi-object clustering

Technical field

The invention belongs to technical field of image processing, relate to the method for image segmentation, can be applicable to Target Recognition.

Background technology

It is in recent years in a popular research direction in image segmentation field that the intelligence computation technology is applied to image segmentation, mainly comprises neural network, genetic algorithm, colony intelligence algorithm and artificial immune system framework.Image segmentation is that piece image is divided into a plurality of zones or object, is basic technology in the Flame Image Process.From the angle of segmentation result, the process of image segmentation gives a label for exactly each pixel, the classification of this label reaction pixel under in segmentation result.In image partition method, each pixel is represented with its characteristics of image, as long as find the label of these features based on feature, just can realize classification to pixel, thereby reach the purpose of image segmentation, therefore, various clustering algorithms have just become one of method that addresses this problem.

Cluster just is meant in the process that does not have under the situation of training sample a stack features to be divided into several classifications, basic thought based on the image segmentation of cluster is: the point that the pixel mapping in the image is become corresponding feature space, if describe suitable that the characteristic variable of different object differential pressures selects, point in the feature space just can be divided into different zones according to certain measurement criterion, shine upon go back to the original image space, obtain segmentation result.

Coleman and Andrews are used for image segmentation with clustering method the earliest, they select gray-scale value for use and by the more derivative statistics of gray-scale value as characteristic variable, adopt the k Mean Method that these unique points are carried out cluster, thereby obtain the segmentation result of image.Nguyen and Colen are shown as the pedigree structure of two Markov random fields with image table, obtain some simple statistical natures and form proper vectors from each image block, utilize these pieces of FCM cluster then.Pan Yunhe will be used for image segmentation based on the probability clustering algorithm of simulated annealing.

Above-mentioned clustering method is when carrying out image segmentation, often have following two shortcomings: (1) evaluation index is single, only uses an objective function, and promptly the point in each class is to its distances of clustering centers sum minimum, carry out cluster with this, thereby cause segmentation result not accurate enough; (2) to the initialization sensitivity, if produced less relatively the separating of some fitness during initialization at random, final to produce the wrong probability of cutting apart just bigger.

Summary of the invention

The objective of the invention is to overcome the deficiency of above-mentioned prior art, propose a kind of image partition method, make segmentation result have better regional consistance, and keep the integrality of image, to improve the quality of image segmentation based on immune multi-object clustering.

The present invention realizes being based on artificial immune system AIS cutting apart of image, AIS is an adaptive system that derives from biological immunological mechanism, in AIS, problem is regarded antigen, antibody is regarded in separating of problem, and the interaction of antigen and antibody, antibody and antibody and the variation of antibody are the essential characteristics of various AIS.Artificial immune system has abilities such as self-adaptation, self-organization, self study, can solve the insoluble challenge of many classic methods.In recent years, AIS receives more and more researchers and engineering technical personnel's concern, and its research field has related to many fields such as machine learning, computer security, optimization.Wherein, the research that artificial immune system is applied to find the solution multi-objective optimization question has caused a lot of scholars' interest.Clustering problem is converted into optimization problem to two objective functions, only at minimizing in the class apart from this target, the multi-object clustering method is mainly used two objective functions with respect to the single goal cluster: minimize distance and maximization between class distance in the class; This clustering technique is sought optimum solution to two objective functions simultaneously by Multipurpose Optimal Method.It is worthy of note that the optimum solution of multi-objective problem is not single, but a Pareto disaggregation is called the Pareto border in object space.

Use is carried out image segmentation based on the method for immune multi-object clustering, can obtain more precise partition result than image segmentation based on the single goal cluster, can overcome the shortcoming that consistance is poor, the border is in disorder, segmentation result is not accurate enough in conventional segmentation methods zone, thereby improve the quality of image segmentation.

Technical scheme of the present invention is: regard image segmentation problem as the multi-object clustering problem, image is carried out feature extraction, with the characteristics of image that extracts as data to be clustered, immune multi-object optimization method with combining local searching carries out multi-object clustering to data, obtaining one group of approximate Pareto separates, from one group is separated, select optimum solution according to the PBM index at last, reach the purpose of image segmentation as classification results.The specific implementation step is as follows:

(1) treat split image and carry out feature extraction, and utilize the dividing ridge method of control mark symbol that image is carried out just cutting apart, according to first segmentation result, the calculated characteristics average obtains data sample to be clustered;

(2) setting maximum iteration time T is 30, current iteration number of times t=0, antibody population B ₀Scale n _dBe 50, clone's population scale n _cBe 50, active population scale n _aBe 20, the cluster classification is counted K, and the size of K is determined according to image to be split; Dominant population, active population and clone population are initialized as respectively

From data to be clustered, select the one by one body of K sample at random, carry out n as antibody population _dInferior aforesaid operations obtains size and is n _dAntibody population B ₀

(3) antagonist group B _tCarry out following optimization, obtain an approximate Pareto disaggregation;

(3.1) difference calculating antibody group B _tIn two target function values of each antibody, these two objective functions are:

$J=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}D(z_k,x_j)$

$\mathrm{XB}=\frac{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{D}^2(z_k,x_j)}{n\×\underset{i\≠j}{\min }\left|\right|z_i-z_j\left|\right|}$

Wherein n is a data sample number to be clustered, and K is a cluster classification number, and x is the cluster data sample, and z is a cluster centre, and D is the Euclidean distance that data sample arrives its cluster centre;

(3.2) at antibody population B _tIn find non-domination antibody as advantage antibody, duplicate all advantage antibody and form interim advantage antibody population DT _tIf DT _tScale is smaller or equal to n _d, make advantage antibody population D _t=DT _t, otherwise, calculate all individual crowding distance values, and select the bigger preceding n of crowding distance value _dThe individual DT that forms _t

(3.3) from advantage antibody population D _tThe middle active antibodies of selecting is formed active population A _tIf, D _tSmall in active population scale n _a, make A _t=D _tOtherwise, calculate D _tIn all individual crowding distances, select the bigger preceding n of crowding distance _aThe active population A of individual composition _t

(3.4) to active population A _tThe ratio clone operations, the antibody population C after obtaining cloning _t

(3.5) to active population A _tCarry out Local Search, obtain the new antibodies group N behind the Local Search _t

(3.6) the antibody population C behind the Comparative Examples clone _tCarry out the simulation scale-of-two and intersect and the polynomial expression mutation operation antibody population C after obtaining making a variation _t';

(3.7) the antibody population C after will making a variation _t', the new antibodies group N that produces of Local Search _tAnd former dominant population D _tMerge, form antibody population B _t, forward step (3.1) to;

(3.8) when iterations reaches T, the advantage antibody D in the output step (3.2) _T+1As approximate Pareto disaggregation;

(4) the approximate Pareto that obtains from step (3.8) separates to concentrate and chooses the separating as optimum solution of PBM index maximum, with it as cluster result;

(5) each pixel for the treatment of split image according to the cluster result of the first result of cutting apart of step (1) and step (4) is redistributed the class mark, obtains final segmentation result.

The present invention compares with existing image partition method based on cluster and has the following advantages:

When (1) the present invention carries out cluster to the pixel feature of the image that extracts, not only considered distance in the class, also considered between class distance simultaneously,, can access regional consistance better pictures segmentation result by these two indexs are estimated.

(2) initial method of the present invention's employing and population are provided with, and compare with the class methods computing velocity faster.

(3) the present invention is to the optimization method of initial antibodies group use, owing to adopted the method for HCS Local Search strategy and artificial immunity multiple-objection optimization, therefore the Pareto disaggregation that obtains has better diversity, by selecting optimum solution, make the inventive method can both obtain better segmentation effect at different images according to the PBM index.

The simulation experiment result shows that the inventive method is compared with existing k Mean Method and based on the image partition method of GAC, can access more accurate image segmentation result.

Description of drawings

Fig. 1 is a general flow chart of the present invention;

Fig. 2 is the sub-process figure of the present invention to multiple-objection optimization;

Fig. 3 is with the emulation segmentation result figure of the present invention on one 3 class texture image;

Fig. 4 is with the emulation segmentation result figure of the present invention on one 4 class texture image;

Fig. 5 is with the emulation segmentation result figure of the present invention on one 2 class SAR image;

Fig. 6 is with the emulation segmentation result figure of the present invention on one 3 class SAR image.

Embodiment

See figures.1.and.2, specific implementation step of the present invention is as follows:

Step 1, image is carried out feature extraction and carries out just cutting apart with dividing ridge method, at first image is carried out three layers of wavelet decomposition, extract the characteristic quantity of 10 filial generations, obtain wavelet character 10 dimensions of each pixel; Secondly select 4 directions according to gray level co-occurrence matrixes, be respectively 0 °, 45 °, 90 ° and 135 °, extract three second degree statisticses as texture characteristic amount along each direction, three statistics are respectively homogeneity district, angle second moment and correlativity, amount to 12 dimensional feature information; Add obtain previously 10 the dimension wavelet characters form 22 dimensional features altogether; In order to reduce calculated amount, adopt dividing ridge method that it is carried out just cutting apart, obtain piece number much smaller than the image slices vegetarian refreshments, represent the feature of this piece with the average of the feature of all pixels in each piece, obtain data to be clustered.

Step 2, setting maximum iteration time T are 30, current iteration number of times t=0, antibody population B ₀Scale n _dBe 50, clone's population scale n _cBe 50, active population scale n _aBe 20, the cluster classification is counted K, and the size of K is determined according to image to be split; Dominant population, active population and clone population are initialized as respectively From data to be clustered, select the one by one body of K sample at random, carry out n as antibody population _dInferior aforesaid operations obtains size and is n _dAntibody population B ₀

Step 3, antagonist group B _tCarry out following optimization, obtain approximate Pareto disaggregation, realize that the detailed process of this step is as follows:

(3.1) calculating target function and upgrade dominant population; According to cutting apart requirement, two objective functions of use are as follows:

$J=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}D(z_k,x_j)$

$\mathrm{XB}=\frac{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{D}^2(z_k,x_j)}{n\×\underset{i\≠j}{\min }\left|\right|z_i-z_j\left|\right|}$

Wherein n is the data number of band cluster, and K is a cluster classification number, and z is a cluster centre, and D is the Euclidean distance that data point arrives its cluster centre.During calculating target function, the classification under the data point is determined to distances of clustering centers according to it, it is referred to a class at nearest center.

(3.2) upgrade dominant population; Realize by following operation: at B _tIn find non-domination antibody as advantage antibody, duplicate all advantage antibody and form interim advantage antibody population and (be expressed as: DT _T+1); If DT _T+1Small in n _d, make D _T+1=DT _T+1Otherwise, calculate all individual crowding distance values, select the bigger preceding n of crowding distance value _dThe individual DT that forms _T+1The crowding distance of body d ∈ D one by one, can calculate according to following formula:

$I(d,D)=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\frac{I_i(d,D)}{f_i^{\max }-f_i^{\min }}$

With

Be respectively the maximal value and the minimum value of i target in the current population, I _i(d D) is defined as follows:

$I_i(d,D)=\left\{\begin{array}{c}\&infin;,\mathrm{if}{f}_i\left(d\right)=\min \left\{f_i\left(d^{\&prime;}\right)\right|d^{\&prime;}\&Element;D\}\mathrm{or}{f}_i\left(d\right)=\max \{f_i\left(d^{\&prime;}\right)|d^{\&prime;}\&Element;D\}\\ \min \{f_i\left(d^{\&prime;}\right)-f_i\left(d^{\&prime;\&prime;}\right)|d^{\&prime;},d^{\&prime;\&prime;}\&Element;D:f_i\left(d^{\&prime;\&prime;}\right)<f_i\left(d\right)<f_i\left(d^{\&prime;}\right)\},\mathrm{otherwise}\end{array}\right.$

Wherein d is an antibody, and f (d) is the objective function of this antibody, and i is the sequence number of objective function;

(3.3) the advantage antibody population is carried out non-domination neighbour selection operation to form active population A _tNon-domination neighbour selection is meant from advantage antibody selects active antibodies, specifically, if D _tSmall in active population scale n _a, make active population A _t=D _tOtherwise, calculate D _tIn all individual crowding distances, select the bigger preceding n of crowding distance _aThe active population A of individual composition _t

(3.4) to active population A _tThe ratio clone operations obtains population C _t

The ratio clone operations is that the individuality that will have big crowding distance is replicated the number of times q that each individuality is replicated more frequently _iValue be calculated as follows:

N wherein _cBe clone's scale, I is a crowding distance, and A is the antibody population that will carry out clone operations, | A| is the scale of antibody population A,

Finger is got integer to asking numerical value

(3.5) to active population A _tCarry out the HCS Local Search, obtain new antibody population N behind the Local Search _t

(3.5.1) judge search condition,, and satisfy searching probability, then carry out Local Search if current iteration number of times t is 10 integral multiple, otherwise, finish Local Search, and establish

(3.5.2) maximum iteration time N is set _NdBe 5, input antibody x, and in the neighborhood of x, produce a new antibodies new at random;

If (3.5.3) new antibodies new domination antibody x, then the antibody behind the Local Search is Xnew=x+h*v, v=new-x wherein, and h is a step-size in search, the size of h is determined according to the quadratic interpolation method;

If (3.5.4) new antibodies new is arranged by antibody x, then the antibody behind the Local Search is Xnew=new+h*v, v=x-new wherein, and h is a step-size in search, the size of h is determined according to the quadratic interpolation method;

If (3.5.5) new antibodies new and antibody x do not arrange mutually, just produce new antibodies new again at random in the neighborhood of antibody x and search for, forward step (4c) to;

If (3.5.6) iterations reaches N _NdThe time, can only find separating of not arranging mutually, then the antibody behind the Local Search is Xnew=x+h*v _Acc, wherein h is a step-size in search, its size is determined by the quadratic interpolation method, v _AccCalculate according to following formula:

$v_{\mathrm{acc}}=\frac{1}{N_{\mathrm{nd}}}\underset{i=1}{\overset{N_{\mathrm{nd}}}{\mathrm{\&Sigma;}}}\frac{\mathrm{new}-x}{\left|\right|\mathrm{new}-x\left|\right|}$

In the formula, N _NdExpression meeting maximum iteration time;

(3.5.7) to active population A _tIn each antibody all carry out above-mentioned steps (4a)～(4f) operation, the Xnew that obtains forms the antibody population N behind the Local Search _t, and output.

(3.6) the antibody population C behind the Comparative Examples clone _tSimulate scale-of-two and intersect and the polynomial expression mutation operation antibody population C after obtaining making a variation _t';

(3.7) the antibody population C after the merging variation _t', the new antibodies group N that produces of Local Search _tAnd former dominant population D _tForm B _t, forward (3.1) to;

(3.8) when iterations reaches T, the advantage antibody D in the output (3.2) _T+1As the Pareto disaggregation.

Step 4, the individuality optimum individual the most of choosing PBM index maximum from approximate Pareto separates are as cluster result; The PBM index definition is as follows:

$\mathrm{PBM}={(\frac{1}{K}\&times;\frac{E_1}{E_K}\&times;D_K)}^2$

Wherein

K is a cluster classification number, and N is a data number to be clustered, and z is a cluster centre.

The cluster result that result that step 5, the watershed divide that obtains according to step 1 are just cut apart and step 4 obtain carries out mark to each pixel of image, just the class of each piece in the cluster result is marked on each pixel that is assigned in this piece, thereby obtain the class mark of each pixel, obtain the final segmentation result of image thus.

Effect of the present invention can further specify by following experiment.

1, experiment content

Use the inventive method and existing k Mean Method and respectively four width of cloth images are carried out split-run test based on the clustering method of heredity, wherein 2 width of cloth texture images and 2 width of cloth SAR images, texture image is calculated classification accuracy rate, the SAR image is estimated the performance of these methods from aspects such as regional consistance, edge maintenances.

2, experimental result

(1) to the segmentation result of texture image

With this method and existing k Mean Method, the genetic cluster method is cut apart two width of cloth texture images, three kinds of methods are distinguished independent operating 10 times, the average classification accuracy rate of this two width of cloth image is as shown in table 1, from table 1, can find out, than k Mean Method and genetic cluster method, this method has all obtained the highest accuracy for two width of cloth images, illustrate that this method can access desirable segmentation result for Study Of Segmentation Of Textured Images, this method is compared genetic cluster method and k Mean Method, the accuracy of on average cutting apart to image one has improved 0.1453 and 0.2344 respectively, and the accuracy of on average cutting apart of image two has been improved 0.0957 and 0.1859 respectively.

Table 1 texture image is on average cut apart accuracy relatively

	The present invention	Genetic cluster	The k average
				Image one	??0.9720	??0.8267	??0.7376
Image two	??0.9576	??0.8619	??0.7717

To the simulation result of three class texture image images one as shown in Figure 3.Wherein Fig. 3 (a) is image one former figure, Fig. 3 (b) is image one canonical reference figure, the emulation segmentation result figure that Fig. 3 (c) obtains for the inventive method, the emulation segmentation result figure that Fig. 3 (d) obtains for the genetic cluster method, the emulation segmentation result figure that Fig. 3 (e) obtains for the k Mean Method.Find out that by Fig. 3 this method is preserved more complete to the edge detailed information.By contrast, the segmentation result profile discrimination that obtains of genetic cluster method and k Mean Method is bad.

To the simulation result of four class texture image images two as shown in Figure 4.Wherein Fig. 4 (a) is image two former figure, Fig. 4 (b) is image two standard drawings, the emulation segmentation result figure that Fig. 4 (c) obtains for the inventive method, the emulation segmentation result figure that Fig. 4 (d) obtains for the genetic cluster method, the emulation segmentation result figure that Fig. 4 (e) obtains for the k Mean Method.As seen from Figure 4, the edge details that this method is cut apart image two is preserved complete, and genetic cluster method and k Mean Method have then obtained wrong segmentation result.

(2) to the segmentation result of SAR image

With the inventive method and existing k Mean Method, genetic cluster method two width of cloth SAR images are cut apart, to the SAR image graph as shown in Figure 5 as three segmentation result, wherein Fig. 5 (a) is the former figure of image three, the emulation segmentation result figure that Fig. 5 (b) obtains for the present invention, the emulation segmentation result figure that Fig. 5 (c) obtains for the genetic cluster method, the emulation segmentation result figure that Fig. 5 (d) obtains for the k Mean Method.As can be seen from Figure 5, the detailed information that the inventive method obtains is the most clear, and the genetic cluster method is failed to make accurately for the zone of the depression at image three harbours and cut apart, and the marginal information that the k Mean Method obtains is then clear inadequately accurately.

To the SAR image graph as shown in Figure 6 as four segmentation result, wherein Fig. 6 (a) is the former figure of image four, the emulation segmentation result figure that Fig. 6 (b) obtains for the present invention, the emulation segmentation result figure that Fig. 6 (c) obtains for the genetic cluster method, the emulation segmentation result figure that Fig. 6 (d) obtains for the k Mean Method.As can be seen from Figure 6 the advantage of the inventive method is the most obvious, detailed information for bridge in the image four and tree, the inventive method can distinguish that genetic cluster method and k Mean Method are then preserved this detailed information and be not sufficiently complete clearly to this details.

Claims (5)

1. the image partition method based on immune multi-object clustering comprises the steps:

(1) treat split image and carry out feature extraction, and utilize the dividing ridge method of control mark symbol that image is carried out just cutting apart, according to first segmentation result, the calculated characteristics average obtains data sample to be clustered;

(2) setting maximum iteration time T is 30, current iteration number of times t=0, antibody population B ₀Scale n _dBe 50, clone's population scale n _cBe 50, active population scale n _aBe 20, the cluster classification is counted K, and the size of K is determined according to image to be split; Dominant population, active population and clone population are initialized as respectively

From data to be clustered, select the one by one body of K sample at random, carry out n as antibody population _dInferior aforesaid operations obtains size and is n _dAntibody population B ₀

(3) antagonist group B _tCarry out following optimization, obtain an approximate Pareto disaggregation;

(3.1) difference calculating antibody group B _tIn two target function values of each antibody, these two objective functions are:

$J=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}D(z_k,x_j)$

$\mathrm{XB}=\frac{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}{D}^2(z_k,x_j)}{n\&times;\underset{i\&NotEqual;j}{\min }\left|\right|z_i-z_j\left|\right|}$

Wherein n is a data sample number to be clustered, and K is a cluster classification number, and x is the cluster data sample, and z is a cluster centre, and D is the Euclidean distance that data sample arrives its cluster centre;

(3.2) at antibody population B _tIn find non-domination antibody as advantage antibody, duplicate all advantage antibody and form interim advantage antibody population DT _tIf DT _tScale is smaller or equal to n _d, make advantage antibody population D _t=DT _t, otherwise, calculate all individual crowding distance values, and select the bigger preceding n of crowding distance value _dThe individual DT that forms _t

(3.3) from advantage antibody population D _tThe middle active antibodies of selecting is formed active population A _tIf, D _tSmall in active population scale n _a, make A _t=D _tOtherwise, calculate D _tIn all individual crowding distances, select the bigger preceding n of crowding distance _aThe active population A of individual composition _t

(3.4) to active population A _tThe ratio clone operations, the antibody population C after obtaining cloning _t

(3.5) to active population A _tCarry out Local Search, obtain the new antibodies group N behind the Local Search _t

(3.6) the antibody population C behind the Comparative Examples clone _tCarry out the simulation scale-of-two and intersect and the polynomial expression mutation operation antibody population C after obtaining making a variation _t';

(3.7) the antibody population C after will making a variation _t', the new antibodies group N that produces of Local Search _tAnd former dominant population D _tMerge, form antibody population B _t, forward step (3.1) to;

(3.8) when iterations reaches T, the advantage antibody D in the output step (3.2) _T+1As approximate Pareto disaggregation;

(4) the approximate Pareto that obtains from step (3.8) separates to concentrate and chooses the separating as optimum solution of PBM index maximum, with it as cluster result;

(5) each pixel for the treatment of split image according to the cluster result of the first result of cutting apart of step (1) and step (4) is redistributed the class mark, obtains final segmentation result.

2. the image partition method based on immune multi-object clustering according to claim 1, the said crowding distance of step (3.2) wherein, calculate according to following formula:

$I(d,D)=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\frac{I_i(d,D)}{f_1^{\max }-f_i^{\min }}$

I is the sequence number of objective function, and k is the number of objective function,

With

Be respectively the maximal value and the minimum value of i target in the current population, I _i(d D) is defined as follows:

$I_i(d,D)=\left\{\begin{array}{c}\&infin;,\mathrm{if}{f}_i\left(d\right)=\min \left\{f_i\left(d^{\&prime;}\right)\right|d^{\&prime;}\&Element;D\}\mathrm{or}{f}_i\left(d\right)=\max \{f_i\left(d^{\&prime;}\right)|d^{\&prime;}\&Element;D\}\\ \min \{f_i\left(d^{\&prime;}\right)-f_i\left(d^{\&prime;\&prime;}\right)|d^{\&prime;},d^{\&prime;\&prime;}\&Element;D:f_i\left(d^{\&prime;\&prime;}\right)<f_i\left(d\right)<f_i\left(d^{\&prime;}\right)\},\mathrm{otherwise}\end{array}\right.$

Wherein d is an antibody, and f (d) is the objective function of this antibody, and i is the sequence number of objective function;

3. the image partition method based on immune multi-object clustering according to claim 1, wherein step (3.4) is described to active population A _tThe ratio clone operations is with A _tIn have the more number of times of individual replicate of big crowding distance, the number of times q that each antibody is replicated _iBe calculated as follows:

N wherein _cBe clone's scale, I is a crowding distance, and A is the antibody population that will carry out clone operations, | A| is the scale of antibody population A,

Finger is got integer to asking numerical value.

4. the image partition method based on immune multi-object clustering according to claim 1, wherein step (3.5) is described to active population A _tCarry out Local Search, comprise the steps:

(4a) judge search condition,, and satisfy searching probability, then carry out Local Search if current iteration number of times t is 10 integral multiple, otherwise, finish Local Search, and establish

(4b) maximum iteration time N is set _NdBe 5, input antibody x, and in the neighborhood of x, produce a new antibodies new at random;

If (4c) new antibodies new domination antibody x, then the antibody behind the Local Search is Xnew=x+h*v, v=new-x wherein, and h is a step-size in search, the size of h is determined according to the quadratic interpolation method;

If (4d) new antibodies new is arranged by antibody x, then the antibody behind the Local Search is Xnew=new+h*v, v=x-new wherein, and h is a step-size in search, the size of h is determined according to the quadratic interpolation method;

If (4e) new antibodies new and antibody x do not arrange mutually, just produce new antibodies new again at random in the neighborhood of antibody x and search for, forward step (4c) to;

If (4f) iterations reaches N _NdThe time, can only find separating of not arranging mutually, then the antibody behind the Local Search is Xnew=x+h*v _Acc, wherein h is a step-size in search, its size is determined by the quadratic interpolation method, v _AccCalculate according to following formula:

$v_{\mathrm{acc}}=\frac{1}{D_{\mathrm{nd}}}\underset{i=1}{\overset{N_{\mathrm{nd}}}{\mathrm{\&Sigma;}}}\frac{\mathrm{new}-x}{\left|\right|\mathrm{new}-x\left|\right|}$

N _NdBe maximum iteration time;

(4g) to active population A _tIn each antibody all carry out above-mentioned steps (4a)～(4f) operation, the Xnew that obtains forms the antibody population N behind the Local Search _t, and output.

5. the image partition method based on immune multi-object clustering according to claim 1, wherein the described approximate Pareto that obtains from step (3.8) of step (4) separates to concentrate and chooses the separating as optimum solution of PBM index maximum, is to calculate by following formula:

$\mathrm{PBM}={(\frac{1}{K}\&times;\frac{E_1}{E_K}\&times;D_K)}^2$

Wherein

K is a cluster classification number, and n is a data number to be clustered, and x is a cluster data, and z is a cluster centre.

Images