CN102930291B - Automatic K adjacent local search heredity clustering method for graphic image - Google Patents

Abstract

The invention discloses an automatic k adjacent local search heredity clustering method for a graphic image, which mainly overcomes the defect of the conventional automatic clustering algorithm that the local optimization is easy to cause. The automatic k adjacent local search heredity clustering method comprises the realization steps of: (1) detecting an outline of an image by utilizing a canny edge detector; (2) describing the outline of the image by utilizing a shape context method and calculating a matched cost matrix of an outline point; (3) matching the outline point by utilizing a dynamic programming method according to the matched cost matrix; (4) converting the matched outline point by utilizing a procrustes analysis method; (5) representing the converted matched outline point and measuring an edit distance between character strings; (6) calculating the distance between the images according to the edit distance of the character strings; (7) clustering the images by utilizing a heredity automatic clustering method; and (8) carrying out k adjacent local search on groups of a heredity method. The automatic k adjacent local search heredity clustering method for the graphic image has the advantages that overall optimization is easy to achieve and an accurate clustering number can be found out.

Description

For the local search heredity automatic clustering method of graph image

Technical field

The invention belongs to technical field of image processing, can be used for the automatic cluster to the obvious image of shape facility.

Background technology

The development of science and technology, impels the research object of people to become increasingly complex.Vision is as one of human knowledge and the Main Means observing the world, and its information is enormous amount but also more complicated not only.The spatial relationship of color, shape, texture and things is all etc. the inscape of visual information, can as the starting point of research.Shape is one of most essential characteristic of having under visually-perceptible meaning of object, the research of this respect has become the importance of computer vision and pattern-recognition, and utilizes shape facility to identify things and one of the main tool that also becomes computer vision and pattern identification research of classifying.Shape extracting, form fit etc. is had to the research of shape.Shape extracting extracts shape information exactly from image, and the more famous method of this respect has canny etc., and form fit can be divided into two aspects such as description, coupling.The method having various shape to describe in recent years and to mate is suggested.Shape description method has Shape-based interpolation edge contour, have Shape-based interpolation region etc.These methods all respectively have relative merits.Form fit weighs the similarity between shape according to certain measurement criterion.Man-to-man form fit is without any prior imformation, often can not distinguish important in shape difference that is insignificant, in any case design more excellent shape descriptor, all cannot address this problem, this just needs to design the internal relation that better algorithm goes for similar shape in shape similarity metric space.

Cluster analysis, as one of the important research direction of data mining, its objective is and allow people's identification data concentrate intensive and sparse region, finds the global distributed schema of data, and the mutual relationship between data attribute.Cluster analysis is applied to every field widely, comprises data analysis, image procossing and market analysis, data mining, statistics, machine learning, spatial database technology, biology and the marketing etc.Traditional clustering algorithm, be in advance given cluster numbers, but be unpredictable by the very multidata cluster numbers of many factors in actual life, the during this time effect of automatic cluster highlights, many scholars propose various method, attempt to solve automatic cluster problem.But these methods often also exist defect: algorithm usually can be absorbed in local optimum, can not get correct cluster numbers.Someone utilizes the thought of Gradient Descent to solve automatic cluster problem, and uses it for the cluster of graph image, but this method still also exists above-mentioned said defect, and algorithm is easily absorbed in local optimum, can not obtain correct cluster numbers.

Summary of the invention

The object of the invention is to for above-mentioned existing technical deficiency, propose a kind of local search heredity automatic clustering method for graph image, to improve the Clustering Effect of image.

Technical thought of the present invention is: by the profile of canny operator detected image, by the profile information of shape context method Description Image, by dynamic programming method, procrustes analysis method and the edit distance distance matrix in conjunction with calculating chart image set, the automatic cluster of image is realized by local search heredity automatic cluster algorithm.Its specific implementation step is as follows:

(1) N width image S to be clustered is inputted _i, i=1 ..., N, utilizes canny operator to carry out rim detection to every width image respectively, obtains the contour images I of each image to be clustered _i;

(2) to each contour images I _icarry out uniform sampling along its outline line, and represent the point P sampling and obtain with Cartesian coordinates _ij=(x _ij, y _ij), i=1 ..., N, j=1 ..., M, wherein, P _ijrepresent a jth point of the i-th width contour images obtained of sampling, x _ij, y _ijbe respectively point P _ijthe transverse and longitudinal coordinate of position, M is the number of sampled point;

(3) distance between any two width images to be clustered is calculated:

(3a) by shape context method, contour images I is described _i, obtain contour images I _ithe Neighbor Points position distribution histogram of each point, and this histogram to be normalized;

(3b) by normalized Neighbor Points position distribution histogram, χ is used ²statistical method, calculates contour images I _ithe Matching power flow matrix of point;

(3c) according to Matching power flow matrix, by dynamic programming method, remove the outlier of two width contour images, obtain the Corresponding matching point of contour images;

(3d) to obtained Corresponding matching point, convert by procrustes analysis method;

(3e) the Corresponding matching point after conversion is converted to character string;

(3f) two image S to be clustered are obtained through step (3a) to (3e) _i, S _jcharacter string;

(3g) to the image S to be clustered obtained _iand S _jcharacter string, with the Edit distance E (S between these two character strings of Edit range observation _i, S _j), then calculate image S to be clustered _iand S _jbetween distance:

d(S _i,S _j)=λ ^u×E(S _i,S _j)/L，

Wherein λ=1.03 are constant, and u is image S to be clustered _ithe outlier number of contour images, L is image S to be clustered _ithe length of character string;

(3h) to the distance d (S of the image to be clustered obtained _i, S _j), be normalized

(4) evolutionary generation t=0 is made, initialization population:

(4a) random generation ρ is more than or equal to the integer k that 2 are less than or equal to ζ _i, i=1 ..., ρ, as the chromosomal cluster numbers of ρ bar, ρ is population scale, and ζ=15 are maximum cluster numbers;

(4b) every bar chromosome is produced at random N number ofly be more than or equal to 1 and be less than or equal to k _iinteger c _j, j=1 ..., N, as this each gene position B chromosomal _jvalue, form this chromosome H _i;

(5) all chromosomal fitness value f are calculated _i, find out the maximum chromosome of fitness value as current optimum dyeing body H, wherein:

f _i=1/θ _i

${\mathrm{\θ}}_i=\mathrm{\α}\×k_i+\frac{1}{k_i}\×{\mathrm{\Σ}}_{s=1}^{k_i}{\mathrm{\ω}}_s$

${\mathrm{\&omega;}}_s=\frac{2}{n_s}\underset{S_k,S_h,S_h\&Element;C_s,k<h}{\mathrm{\&Sigma;}}d{(S_k,S_h)}^2$

Wherein, α=0.5 is controling parameters, k _iby being calculated chromosomal cluster numbers, d (S _k, S _h) be image S to be clustered _kand S _hdistance, n _sthe secondary number of image to be clustered contained by s class cluster image, C _sbe the set of s class image to be clustered, S _kand S _hit is the image to be clustered in s class image to be clustered;

(6) according to obtained chromosomal fitness value f _i, with roulette wheel selection to all chromosome H _iselect, select to carry out genetic manipulation centre for chromosome H ' _i;

(7) to the centre obtained for chromosome H ' _i, with probability σ ₁, σ ₁∈ [0,1] carries out single-point interlace operation;

(8) to centres all after single-point interlace operation for chromosome H " _i, with probability σ ₂, σ ₂∈ [0,1] carries out single-point mutation operation;

(9) calculate after single-point interlace operation all centres for chromosome H " _ifitness value f _i;

(10) middle for chromosome H to every bar " _i, produce random number r, a r ∈ [0,1], if r<0.5, perform step (11), otherwise perform step (12);

(11) to centre for chromosome H " _ieach gene position B _j, produce random number ψ, ψ ≠ c _j, 1≤ψ≤k _i, k _ifor centre is for chromosome H " _icluster numbers, c _jfor gene position B _jvalue, by gene position B _jvalue c _jchange ψ into, calculate now for chromosome H " ' _ifitness value f _i', and with change before centre for chromosome fitness value f _irelatively, if f ' _i>f _i, then gene position B _jvalue be set to ψ, perform step (13), otherwise gene position B _jvalue be set to c _j, perform step (13);

(12) to centre for chromosome H " _icarry out k neighbour Local Search:

(12a) by centre for chromosome H " _ieach gene position B _jcorresponding image S to be clustered _j, with other image S to be clustered _ldistance d (S _j, S _l) sort from small to large;

(12b) from sorted distance, front K=5 is chosen apart from corresponding image to be clustered, as K secondary arest neighbors image S ' _v;

(12c) to selected K secondary arest neighbors image S ' _vcorresponding gene position B _v, the number of statistics different genes place value, the gene place value τ finding number maximum, if τ=c _j, c _jfor image S to be clustered _jcorresponding gene position B _jvalue, then perform step (13);

(12d) by centre for chromosome H " _igene position B _jvalue change τ into, calculate now in the middle of for chromosome H " ' _ifitness value f " _i, for chromosome H in the middle of before changing " _ifitness value f _iif, f " _i>f _i, then gene position B _jvalue be set to τ, otherwise gene position B _jvalue be set to c _j;

(13) more all centres are for chromosomal fitness value f _ifind out the dyeing that fitness value is maximum, this chromosome is compared with current optimum dyeing body H, if this chromosomal fitness value is greater than the fitness value of current optimum dyeing body H, then by this chromosome replication to current optimum dyeing body H, if the fitness value of current optimum dyeing body H is greater than this chromosomal fitness value, then current optimum dyeing body H is copied to this chromosome;

(14) 1 is added to evolutionary generation t value, if now t is less than maximum evolutionary generation T=100, then return step (6), otherwise, by each gene position B of current optimum dyeing body H _ivalue c _imake gene position B _icorresponding image S to be clustered _iclass mark, image to be clustered identical for all class marks is divided into a class, the cluster numbers of the image to be clustered after output category and current optimum dyeing body H.

The present invention compared with prior art has the following advantages:

1, the present invention owing to adding local searching operator after genetic manipulation, therefore can adjust the evolution process of population among a small circle, adds the diversity of population, be conducive to overcoming the defect being absorbed in local optimum, improve the Clustering Effect of image.

2, the present invention carries out Local Search due to the neighbor information according to image, improves the direction of Evolution of Population, accelerates speed of convergence.

Accompanying drawing explanation

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

Fig. 2 is the wherein 4 sub-picture examples in test pattern image set kimia216;

Fig. 3 is with the cluster simulated effect figure of the present invention to 30 sub-pictures in test pattern image set kimia216;

Fig. 4 is to the cluster simulated effect figure of 30 sub-pictures in test pattern image set kimia216 with contrast algorithm.

Embodiment

With reference to Fig. 1, realization of the present invention comprises the steps:

Step one: input N width image S to be clustered _i, i=1 ..., N, utilizes canny operator to carry out rim detection to every width image respectively, obtains the contour images I of each image to be clustered _i, in this example, N gets 30 but is not limited to 30, canny operator and proposed in 1986 years by John Canny.

Step 2: to each contour images I _icarry out uniform sampling along its outline line, and represent the point P sampling and obtain with Cartesian coordinates _ij=(x _ij, y _ij), i=1 ..., N, j=1 ..., M, wherein, P _ijrepresent a jth point of the i-th width contour images obtained of sampling, x _ij, y _ijbe respectively point P _ijthe transverse and longitudinal coordinate of position, M=100 is the number of sampled point.

Step 3: calculate the distance between any two width images to be clustered.

3.1) by shape context method, contour images I is described _i, obtain contour images I _ithe Neighbor Points position distribution histogram of each point, and this histogram to be normalized:

3.1.1) to contour images I _i, calculate the ultimate range between all point

d=max _{1≤m＜n≤M}{||P _im-P _in||}

$\left|\right|P_{\mathrm{im}}-P_{\mathrm{in}}\left|\right|=\sqrt{{(x_{\mathrm{im}}-x_{\mathrm{in}})}^2+{(y_{\mathrm{im}}-y_{\mathrm{in}})}^2},$

Ultimate range d is converted to logarithm log (d), and this logarithmic mean is quantified as b ₁individual grade, is quantified as b by average for 360 ° of angles ₂individual grade, with point P _ijfor initial point, by this b ₁× b ₂individual grade classification goes out b ₁× b ₂individual region;

3.1.2) to contour images I _ieach point P _ij, by all point P except this point _ikpolar coordinates P ' is converted to by Cartesian coordinates _ik=(r _ik, θ _ik), then be converted to log-polar P " _ik=(r ' _ik, θ ' _ik), wherein:

$r_{\mathrm{ik}}=\sqrt{{(x_{\mathrm{ik}}-x_{\mathrm{ij}})}^2+{(y_{\mathrm{ik}}-y_{\mathrm{ij}})}^2}$

${\mathrm{\&theta;}}_{\mathrm{ik}}=\arctan \left(\frac{y_{\mathrm{ik}}-y_{\mathrm{ij}}}{x_{\mathrm{ik}}-x_{\mathrm{ij}}}\right)$

r′ _ik=log(r _ik)

θ′ _ik=θ _ik-θ _O,

${\mathrm{\&theta;}}_0=\arctan \left(\frac{y_0-y_{\mathrm{ij}}}{x_0-x_{\mathrm{ij}}}\right)$

$x_0=\frac{1}{M}\&times;\underset{l=1}{\overset{M}{\mathrm{\&Sigma;}}}{x}_{\mathrm{il}}$

$y_0=\frac{1}{M}\&times;\underset{l=1}{\overset{M}{\mathrm{\&Sigma;}}}{y}_{\mathrm{il}}$

In log-polar system, statistics falls into ready-portioned b ₁× b ₂point P in the regional in individual region " _iknumber c _l, l=1 ..., b ₁× b ₂, recycling experience density formula is normalized, and obtains point P _ijneighbor Points position distribution normalization histogram h _j(s):

$h_j\left(s\right)=\frac{c_s}{\underset{l=1}{\overset{b_1\&times;b_2}{\mathrm{\&Sigma;}}}{c}_l},$ s=1,…,b ₁×b ₂；

3.2) to the Neighbor Points position distribution normalization histogram obtained, according to formulae discovery Matching power flow Matrix C (i, j):

$C(i,j)=\frac{1}{2}\underset{s=1}{\overset{b_1\&times;b_2}{\mathrm{\&Sigma;}}}\frac{{[h_i\left(s\right)-h_j\left(s\right)]}^2}{h_i\left(s\right)+h_j\left(s\right)};$

3.3) according to the Matching power flow matrix obtained, by dynamic programming method, remove the outlier of two width contour images, obtain the Corresponding matching point of contour images, the concrete implementation step of dynamic programming method is shown in that M.R.Daliri and V. Torre was published in the article Robustsymbolic representation for shape recognition and retrieval on Pattern Recognition in 2008.

3.4) to obtained Corresponding matching point, convert by procrustes analysis method, the concrete implementation step of procrustes analsis method is shown in that Mohammad Reza Daliri and Vincent Torre was published in the article Shape and texture clustering:Best estimate for the clustersnumber on Imageand Vision Computing in 2009.

3.5) the Corresponding matching point after conversion is converted to character string, its concrete implementation step is shown in that Mohammad RezaDaliri and Vincent Torre was published in the article Shape andtexture clustering:Best estimate for the clusters number on Image and Vision Computing in 2009.

3.6) through step 3.1) to 3.5) obtain two image S to be clustered _i, S _jcharacter string.

3.7) to the image S to be clustered obtained _iand S _jcharacter string, with the Edit distance E (S between these two character strings of Edit range observation _i, S _j), then calculate image S to be clustered _iand S _jbetween distance:

d(S _i,S _j)=λ ^u×E(S _i,S _j)/L，

Wherein λ=1.03 are constant, and u is image S to be clustered _ithe outlier number of contour images, L is image S to be clustered _ithe length of character string.

The measuring process of Edit distance is shown in the article Learning string edit distance that Eric SvenRistad and PeterN.Yianilos delivered on IEEE Trans.PAMI in 1998.

3.8) to the distance d (S of the image to be clustered obtained _i, S _j), be normalized.

Step 4: make evolutionary generation t=0, initialization population:

4.1) random generation ρ is more than or equal to the integer k that 2 are less than or equal to ζ _i, i=1 ..., ρ, as the chromosomal cluster numbers of ρ bar, ρ is population scale, and ζ=15 are maximum cluster numbers.

4.2) every bar chromosome is produced at random N number ofly be more than or equal to 1 and be less than or equal to k _iinteger c _j, j=1 ..., N, as this each gene position B chromosomal _jvalue, form this chromosome H _i.

Step 5: calculate all chromosomal fitness value f _i, find out the maximum chromosome of fitness value as current optimum dyeing body H, wherein:

f _i=1/θ _i

${\mathrm{\&theta;}}_i=\mathrm{\&alpha;}\&times;k_i+\frac{1}{k_i}{\mathrm{\&Sigma;}}_{s=1}^{k_i}{\mathrm{\&omega;}}_s$

${\mathrm{\&omega;}}_s=\frac{2}{n_s}\underset{S_k,S_h\&Element;C_s,k<h}{\mathrm{\&Sigma;}}d{(S_k,S_h)}^2$

Wherein, α=0.5 is controling parameters, k _iby being calculated chromosomal cluster numbers, d (S _k, S _h) be image S to be clustered _kand S _hdistance, n _sthe secondary number of image to be clustered contained by s class cluster image, C _sbe the set of s class image to be clustered, S _kand S _hit is the image to be clustered in s class image to be clustered.

Step 6: according to obtained chromosomal fitness value f _i, to all chromosome H _iselect, select to carry out genetic manipulation centre for chromosome H ' _i.

6.1) according to obtained chromosomal fitness value f _i, pass through formula

$F_i=\frac{{\mathrm{\&Sigma;}}_{s=1}^i{f}_s}{{\mathrm{\&Sigma;}}_{j=1}^N{f}_j},$

Calculate all chromosomal accumulated probability F _i;

6.2) random generation ρ number γ _j, γ _j∈ [0,1], to each random number γ _j, find subscript value l, make F _l-1< γ _j≤ F _lset up, by l chromosome H _l, copy to middle for chromosome H ' _j.

Step 7: to the centre obtained for chromosome, with probability σ ₁carry out single-point interlace operation, σ ₁∈ [0,1].

The concrete steps of single-point interlace operation are shown in that Ujjwal Maulik and Sanghamitra Bandyopadhyay is published in the article Genetic algorithm-based clustering technique on Pattern Recognition.

Step 8: to centres all after single-point interlace operation for chromosome H " _i, with probability σ ₂, σ ₂∈ [0,1] carries out single-point mutation operation.

The concrete steps of single-point mutation operation are shown in that Ujjwal Maulik and Sanghamitra Bandyopadhyay is published in the article Genetic algorithm-based clustering technique on Pattern Recognition.

Step 9: to calculate after single-point interlace operation all centres for chromosome H " _ifitness value f _i.

Step 10: middle for chromosome H to every bar " _i, produce random number r, a r ∈ [0,1], if r<0.5, perform step 11, otherwise perform step 12.

Step 11: to centre for chromosome H " _ieach gene position B _j, produce random number ψ, ψ ≠ c _j, 1≤ψ≤k _i, k _ifor centre is for chromosome H " _icluster numbers, c _jfor gene position B _jvalue, by gene position B _jvalue c _jchange ψ into, calculate now for chromosome H " ' _ifitness value f ' _i, and with change before centre for chromosome fitness value f _irelatively, if f ' _i>f _i, then gene position B _jvalue be set to ψ, perform step 13, otherwise gene position B _jvalue be set to c _j, perform step 13.

Step 12: to centre for chromosome H " _icarry out k neighbour Local Search:

12.1) by centre for chromosome H " _ieach gene position B _jcorresponding image S to be clustered _jwith other image S to be clustered _ldistance d (S _j, S _l), sort from small to large;

12.2) from sorted distance, front K=5 is chosen apart from corresponding image to be clustered, as K secondary arest neighbors image S ' _v;

12.3) to selected K secondary arest neighbors image S ' _vcorresponding gene position B _v, the number of statistics different genes place value, the gene place value τ finding number maximum, if τ=c _j, c _jfor image S to be clustered _jcorresponding gene position B _jvalue, then perform step rapid 13;

12.4) by centre for chromosome H " _igene position B _jvalue change τ into, calculate now in the middle of for chromosome H " ' _ifitness value f " _i, for chromosome H in the middle of before changing " _ifitness value f _icompare, if f " _i>f _i, then gene position B _jvalue be set to τ, otherwise gene position B _jvalue be set to c _j.

Step 13: more all centres are for chromosomal fitness value f _ifind out the dyeing that fitness value is maximum, this chromosome is compared with current optimum dyeing body H, if this chromosomal fitness value is greater than the fitness value of current optimum dyeing body H, then by this chromosome replication to current optimum dyeing body H, if the fitness value of current optimum dyeing body H is greater than this chromosomal fitness value, then current optimum dyeing body H is copied to this chromosome.

Step 14: add 1 to evolutionary generation t value, now, if t is less than maximum evolutionary generation T=100, then returns step (6), otherwise, by each gene position B of current optimum dyeing body H _ivalue c _imake gene position B _icorresponding image S to be clustered _iclass mark, image to be clustered identical for all class marks is divided into a class, the cluster numbers of the image to be clustered after output category and current optimum dyeing body H.

Effect of the present invention can be further illustrated by following experiment:

1, the image of emulation experiment employing:

The 30 width images that experiment employs in test pattern image set kimia216 are as test pattern, and these images are the obvious bianry image of shape facility, and wherein Fig. 2 is 4 pairs central in test pattern image set kimia216.

2, the optimum configurations condition of emulation experiment:

Setup parameter is: point number of samples M=100, polar coordinates Neighbor Points position distribution histogram diameter quantized value b ₁=5, polar coordinates Neighbor Points position distribution histogram angular quantification value b ₂=12, the penalty value 0.5 of dynamic programming method, Edit distance measure penalty value 0.5, Edit distance measure consecutive point number 5, Population Size ρ=30, maximum cluster numbers ζ=15, the secondary number K=5 of maximum evolutionary generation T=100, k arest neighbors image, crossover probability σ ₁=0.8, mutation probability σ ₂=0.2, controling parameters α=0.1.

3, emulation experiment environment:

Be core2 2.4HZ at CPU, internal memory 2G, WINDOWS XP system uses C++ to emulate.

4, content is emulated

Emulation 1

With the present invention and existing shape image clustering algorithm, independently run 20 times, record the cluster numbers of each result, three indexs such as ARI, MS, calculate its average and standard deviation as shown in table 1.

Table 1 the present invention contrasts with the Validity Index of contrast algorithm

As can be seen from Table 1, under three kinds of Validity Indexes, performance of the present invention is all better than contrast algorithm, demonstrates validity of the present invention.

Emulation 2

Carry out cluster with the present invention to 30 sub-pictures in test pattern image set kimia216, obtain the image of cluster result as being positioned at same row in Fig. 3, Fig. 3 and belong to same class, the image being positioned at different rows belongs to inhomogeneity.

Emulation 3

Carry out cluster with contrast algorithm to 30 sub-pictures in test pattern image set kimia216, obtain the image of cluster result as being positioned at same row in Fig. 4, Fig. 4 and belong to same class, the image being positioned at different rows belongs to inhomogeneity.

Can be found out by above-mentioned emulation experiment, the present invention can obtain good Clustering Effect to test pattern.Relatively going up numerically, and relatively the going up of visual effect, verify rationality of the present invention and validity all effectively.

See that Mohammad Reza Daliri and Vincent Torre to be published in the article Shape and texture clustering:Bestestimate for the clustersnumber on Image and Vision Computing in 2009 from the contrast algorithm emulation 1 and emulation 3.

Claims (1)

1., for a local search heredity automatic clustering method for graph image, comprise the steps:

(1) N width image S to be clustered is inputted _p, p=1 ..., N, utilizes canny operator to carry out rim detection to every width image respectively, obtains the contour images I of each image to be clustered _p;

(2) to each contour images I _pcarry out uniform sampling along its outline line, and represent the point P sampling and obtain with Cartesian coordinates _pq=(x _pq, y _pq), p=1 ..., N, q=1 ..., M, wherein, P _pqrepresent q point of the p width contour images obtained of sampling, x _pq, y _pqbe respectively point P _pqthe transverse and longitudinal coordinate of position, M is the number of sampled point;

(3) distance between any two width images to be clustered is calculated;

(3a) by shape context method, contour images I is described _p1and I _p2, obtain the Neighbor Points position distribution histogram of each point of contour images, and histogram be normalized;

(3b) by normalized Neighbor Points position distribution histogram, χ is used ²statistical method, calculates the Matching power flow matrix of point on contour images;

(3c) according to Matching power flow matrix, by dynamic programming method, remove the outlier of two width contour images, obtain the Corresponding matching point of contour images;

(3d) to obtained Corresponding matching point, convert by procrustes analysis method;

(3e) the Corresponding matching point after conversion is converted to character string;

(3f) obtain the character string S of two images to be clustered to (3e) through step (3a) _p1and S _p2;

(3g) according to the character string of the image to be clustered obtained, the distance between image to be clustered is calculated:

d(S _p1,S _p2)＝λ ^u×E(S _p1,S _p2)/L，

Wherein E (S _p1, S _p2) for image to be clustered character string between Edit distance, λ=1.03 are constant, and u is the outlier number of the contour images of image to be clustered, and L is the length of the character string of image to be clustered;

(3h) to the distance between the image all to be clustered obtained, be normalized;

(4) evolutionary generation t=0 is made, initialization population:

(4a) random generation ρ is more than or equal to the integer k that 2 are less than or equal to ζ _i, i=1 ..., ρ, as the chromosomal cluster numbers of ρ bar, ρ=30 are population scale, and ζ=15 are maximum cluster numbers;

(4b) every bar chromosome is produced at random N number ofly be more than or equal to 1 and be less than or equal to k _iinteger c _j, j=1 ..., N, as this each gene position B chromosomal _jvalue, form this chromosome H _i;

(5) all chromosomal fitness value f are calculated _i ⁰, find out the maximum chromosome of fitness value as current optimum dyeing body H;

Fitness value f _i ⁰, by following formulae discovery:

f _i ⁰＝1/θ _i

${\mathrm{\&theta;}}_i=\mathrm{\&alpha;}\&times;k_i+\frac{1}{k_i}\&times;{\mathrm{\&Sigma;}}_{s=1}^{k_i}{\mathrm{\&omega;}}_s$

${\mathrm{\&omega;}}_s=\frac{2}{n_s}\underset{S_k,S_h\&Element;C_s,k<h}{\mathrm{\&Sigma;}}d{(S_k,S_h)}^2$

Wherein, α=0.5 is controling parameters, k _iby being calculated chromosomal cluster numbers, d (S _k, S _h) be image S to be clustered _kand S _hdistance, n _sthe secondary number of image to be clustered contained by s class cluster image, C _sbe the set of s class image to be clustered, S _kand S _hit is the image to be clustered in s class image to be clustered;

(6) according to obtained chromosomal fitness value f _i ⁰, with roulette wheel selection to all chromosome H _iselect, select to carry out genetic manipulation centre for chromosome H ' _i;

(7) to the centre obtained for chromosome H ' _i, with probability σ ₁, σ ₁∈ [0,1] carries out single-point interlace operation;

(8) to centres all after single-point interlace operation for chromosome H " _i, with probability σ ₂, σ ₂∈ [0,1] carries out single-point mutation operation, obtains the centre after single-point mutation operation for chromosome H _i" ';

(9) use step (5) method calculate after single-point mutation operation all centres for chromosome H _i" ' fitness value f _i;

(10) middle for chromosome H to every bar _i" ', produces random number r, a r ∈ [0,1], if r<0.5, performs step (11), otherwise performs step (12);

(11) to centre for chromosome H _i" ' each gene position B _j, produce random number ψ, ψ ≠ c _j, 1≤ψ≤k _i', k _i' be middle for chromosome H _i" ' cluster numbers, c _jfor gene position B _jvalue, by gene position B _jvalue c _jchange ψ into, calculate now for chromosome H _i" ' fitness value f ' _i, and with change before centre for chromosome fitness value f _irelatively, if f ' _i>f _i, then gene position B _jvalue be set to ψ, perform step (13), otherwise gene position B _jvalue be set to c _j, perform step (13);

(12) to centre for chromosome H _i" ' carry out k neighbour Local Search:

(12a) by centre for chromosome H _i" ' each gene position B _jcorresponding image S to be clustered _j, with other image S to be clustered _ldistance d (S _j, S _l) sort from small to large;

(12b) from sorted distance, front K=5 is chosen apart from corresponding image to be clustered, as K secondary arest neighbors image S ' _v;

(12c) to selected K secondary arest neighbors image S' _vcorresponding gene position B _v, the number of statistics different genes place value, the gene place value τ finding number maximum, if τ=c _j, c _jfor image S to be clustered _jcorresponding gene position B _jvalue, perform step (13);

(12d) by centre for chromosome H _i" ' gene position B _jvalue change τ into, calculate now in the middle of for chromosome H _i" ' fitness value f _i", by front to itself and change middle for chromosome H _i" ' fitness value f _icompare, if f _i" >f _i, then gene position B _jvalue be set to τ, otherwise gene position B _jvalue be set to c _j;

(13) more all centres are for chromosome H _i" ' fitness value f _i"; and find out the maximum chromosome of fitness value; compared with current optimum dyeing body H by this chromosome; if this chromosomal fitness value is greater than the fitness value of current optimum dyeing body H; then by this chromosome replication to current optimum dyeing body H; if the fitness value of current optimum dyeing body H is greater than this chromosomal fitness value, then current optimum dyeing body H is copied to this chromosome;

(14) 1 is added to evolutionary generation t value, if now t is less than maximum evolutionary generation T=100, then return step (6), otherwise, by each gene position B of current optimum dyeing body H _jvalue c _jmake gene position B _jcorresponding image S to be clustered _jclass mark, image to be clustered identical for all class marks is divided into a class, the cluster numbers of the image to be clustered after output category and current optimum dyeing body H.