CN101299243A - Method of image segmentation based on immune spectrum clustering - Google Patents

Abstract

The invention discloses an image segmentation method based on immunity spectrum clustering, which includes: 1. extracting texture characteristic of the input image, representing each pixel point in the image with an eigenvector to obtain a characteristic set; 2. mappings the characteristic set to a linear measure space by spectrum clustering to a mapping set; 3. dividing the category number according to the given image, accidentally selecting the corresponding number of data from the mapping set as the initial clustering center, executing cloning, variation, selection and judgement in sequence, to find out a optimum clustering center with the same category number with the initial clustering center; 4. dividing all pixel points of the characteristic set to an optimum clustering center nearest to the pixel points, and giving each pixel point a category mark according to the category of optimum clustering center where the pixel point locates to complete the image segmentation. Compared with the prior technology, the invention has advantages of insensitivity to initialization, quick convergence to global optimum and high specification accuracy, which can be used in the image segmentation of SAR image processing and computer visual sense field.

Description

Image partition method based on immune spectrum clustering

Technical field

The invention belongs to technical field of image processing, relate to a kind of image partition method, can be used for image segmentation SAR Flame Image Process and computer vision field.

Background technology

Image segmentation is one of technology basic and crucial in Flame Image Process and the computer vision, is meant the zone that image is divided into each tool characteristic, and the zone in the image is meant that interconnect, as to have consistent " meaningful " attribute pixel set.In a lot of machine vision and image processing algorithm, characteristics of image has been made simple hypothesis: the intensity in image local zone is that different zones has different intensity uniformly.But also there is the image of many actual objects not satisfy this hypothesis, the for example surface of timber, hair, braid, meadow, sandy beach, its image has the variation of intensity, this variation formed certain repetition, visually appreciable " texture ", at this moment the zones of different in the image is discerned according to texture rather than shape or intensity homogeneity.Therefore at first to extract image texture features, obtaining the characteristic set of piece image, and then this characteristic set be cut apart.

Can be summed up as a clustering problem the cutting apart of characteristic set to image.Cluster is exactly the process of things being distinguished and being classified according to certain requirement and rule.Cluster analysis then is meant with the method research of mathematics and handles given object, is a kind of of multivariate statistical analysis, also is an important branch of no supervised recognition.It does not have one the sample set of classification mark to be divided into several subclass by certain criterion, make similar sample be classified as a class as far as possible, and dissimilar sample is divided in the different classes as far as possible.Clustering problem is that one of machine learning and area of pattern recognition are relatively more active and have challenging research direction always, has produced a large amount of related algorithms that addresses this problem in recent years.Existing clustering method based on production model is owing to want the operation parameter density Estimation, have to simplify the model of problem, as the distribution of supposing each class is a Gaussian distribution, just make these algorithms only good effect be arranged having on the data of convex structure, be unsuitable for having the clustering problem of the complex form of any distribution.Secondly, because there is local minimum in log-likelihood, thereby when obtaining satisfied cluster, have to attempt several initial configurations.Other algorithm is as the clustering algorithm based on the center, be the representational k-of having mean algorithm, on the data acquisition that hyper-sphere distributes, good performance arranged, be unsuitable for the arbitrary shape cluster, and utilize iterative optimization method to seek optimum solution, can not guarantee to converge to globally optimal solution.

The effective clustering method of a nearest class begins to be subjected to extensive concern.These class methods are based upon on the spectrogram theoretical foundation, utilize the proper vector of data similar matrix to carry out cluster, thereby are referred to as spectral clustering.The spectral clustering algorithm is a kind of method based on the point-to-point transmission similarity relation, is applicable to non-measure space.Because the dimension of this algorithm and data point is irrelevant, and is only relevant with the number of data point, thereby can avoid by the too high singularity problem that causes of the dimension of proper vector.The spectral clustering algorithm is again a discriminant method, need not suppose the global structure of data, represent that belong to of a sort possibility at 2 but at first collect local message, make global decisions according to a certain clustering criteria then, all data points are divided in the different data acquisitions.Usually such criterion can obtain explaining that this embedded space is to be opened by certain several proper vector of data matrix in an embedded space.For example the NJW method is exactly a kind of dividing method to image in the spectral clustering, its process is, at first extract image texture features constitutive characteristic collection, mapping obtains mapping ensemblen to feature set again, use the k-mean cluster to obtain cluster centre to mapping ensemblen then, with the cluster centre that obtains each pixel of image is composed classification at last, the final segmentation result of class mark expression, as shown in Figure 1.

This spectral method the cause for the success is: by feature decomposition, can obtain the globally optimal solution of clustering criteria in unfixed continuous domain.What yet the spectral clustering algorithm of this classics adopted in the territory in mapping is the k-mean cluster, and it is responsive and easily be absorbed in local optimum to initialization that k-average itself exists, and is difficult for the shortcoming carried out, so it is relatively poor to carry out the result of image segmentation with it.

The content of invention

The objective of the invention is to overcome the deficiency of prior art, be that the k-mean cluster that adopts in the Classical Spectrum clustering algorithm has the initialization sensitivity, easily be absorbed in the shortcoming of local optimum, a kind of image partition method based on immune spectrum clustering is proposed, to realize improving the image segmentation effect, this method has the ability of the non-protruding cluster of identification, and is easy in actual applications carry out.

Technical scheme of the present invention is: utilize the dimension reduction characteristic of spectral clustering to obtain the distribution of data at mapping space; On this basis, construct a kind of new immune clone clustering and be used at mapping space sample being carried out cluster, and by the spectral clustering mapping, can obtain the cluster of compacting in the mapping territory, detailed process is as follows:

(1) extracts the textural characteristics of input picture, and each pixel in the image is represented with a proper vector, obtain the characteristics of image collection;

(2) by the spectral clustering algorithm characteristics of image collection is mapped to a linear measurement space, obtains its mapping (enum) data collection;

(3) according to given image division classification number, select the respective counts destination data at random as initial cluster center from mapping (enum) data is concentrated, clone successively, make a variation, select and adjudicate, find out and the same number of optimum cluster centre of its classification;

(4) all pixels with the characteristics of image collection are divided in its nearest optimum cluster centre, and each pixel is composed a class mark by the classification of the optimum cluster centre in its place, are the segmentation result of image.

The implementation procedure of described step (3) is as follows:

(3a) from mapping ensemblen Y, select c data at random, and be expressed as initial cluster center:

$B\left(k\right)=\left[\begin{array}{c}{Y}_1\\ Y_2\\ \·\·\·\\ Y_c\end{array}\right]=\left[\begin{array}{cccc}{y}_{11}& y_{12}& \·\·\·& y_{1c}\\ y_{21}& y_{22}& \·\·\·& y_{2c}\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ y_{c1}& y_{c2}& \·\·\·& y_{\mathrm{cc}}\end{array}\right]$

In the formula, k is an iteration step, k=0 during initialization, Y ₁, Y ₂Y _cThe c that expression is selected from mapping ensemblen Y at random data as initial cluster center;

(3b) initialized cluster centre B (k) is utilized clone operations operator T _c ^CCarry out clone operations, the cluster centre B ' after obtaining cloning (k) is:

$B^{\′}\left(k\right)=T_c^C\left[B\left(k\right)\right]={\left[\begin{array}{cccc}{Y}_1& Y_1& \·\·\·& Y_1\\ Y_2& Y_2& \·\·\·& Y_2\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ Y_c& Y_c& \·\·\·& Y_c\end{array}\right]}_{(c*p_c)}$

In the formula, (c*p _c) size of representing matrix, p _cBe clone's scale;

(3c) B ' (k) is carried out cluster centre B after mutation operation obtains making a variation " (k) be:

B″(k)＝B′(k)±rand*p _m

In the formula, rand represents a random number between [0～1], p _mBe the variation probability;

" (k) carry out cluster centre B (k+1) after selection operation obtains selecting; promptly calculate data among the mapping ensemblen Y (3d) to B to the cluster centre B " quadratic sum of Euclidean distance (k), as the affinity of this cluster centre, from every class cluster centre, select have maximum affinity cluster centre as B (k+1);

(3e) repeating step (3b)-(3d), and according to predefined iteration number iteration step k is judged, if current iteration step greater than predefined iteration number, then stops iteration, otherwise continue to carry out (3b).

The present invention compared with prior art has following advantage:

1. have the global optimum of rapidly converging to, and to the insensitive characteristic of initialization

Immune clone clustering is a kind of emerging artificial immune system method, and it is by the immune antibody cloning Selecting Mechanism and Procedure of biology, and structure is applicable to clone's operator of artificial intelligence.Because the clone selection clustering based on clone's operator is colony's search strategy, is difficult for being absorbed in local minimum in search, finally can obtains the globally optimal solution of problem, and have speed of convergence faster with bigger probability; Immune clone clustering is to determine that search combines with random search, to have the randomness that concurrency and search change that this makes it insensitive to initialization, can begin to search global optimum from initialization arbitrarily simultaneously.

2. replace the k-mean cluster with immune clone clustering

In traditional spectral clustering method, be to adopt the k-average that data are carried out cluster in mapping space as the NJW method, yet there are some intrinsic shortcomings in k-average itself, makes the cluster result of spectral clustering be affected.The present invention combines spectral clustering and immune clone clustering and has constructed an immune spectrum clustering method, has overcome the deficiency of k-average.

3. segmentation effect is good, and nicety of grading increases

In order to verify the superiority of the image partition method based on immune spectrum clustering of the present invention, four class data: Iris, Sonar, Wine and Segment from the UCI database, have been selected respectively, from the texture image storehouse, selected two width of cloth texture images and two width of cloth SAR images and assessed the performance of different dividing methods, simultaneously with the present invention and traditional based on the k-mean cluster image partition method and compare based on the image partition method of Classical Spectrum cluster.Emulation experiment shows with the k-mean cluster to be compared with the Classical Spectrum cluster, and immune spectrum clustering method of the present invention has not only reduced intrinsic dimensionality, and has improved segmentation precision, thereby has verified that algorithm is used for the validity of image segmentation.

Description of drawings

Fig. 1 is the process synoptic diagram of existing NJW image partition method;

Fig. 2 is the process synoptic diagram of image partition method of the present invention;

Fig. 3 is with the present invention and the two kinds of control methodss emulation experiment segmentation result figure to two class texture images;

Fig. 4 is with the present invention and the two kinds of control methodss emulation experiment segmentation result figure to four class texture images;

Fig. 5 is with the present invention and the two kinds of control methodss emulation experiment segmentation result figure to two class SAR images;

Fig. 6 is with the present invention and the two kinds of control methodss emulation experiment segmentation result figure to three class SAR images.

Embodiment

With reference to Fig. 2, implementation process of the present invention is as follows:

Step 1 is extracted the textural characteristics of input picture, and each pixel in the image is represented with a proper vector, obtains characteristics of image collection S.

The input piece image at first calculates its gray level co-occurrence matrixes, calculates related coefficient, entropy and the energy feature of each pixel as image texture features by gray level co-occurrence matrixes then; Simultaneously this width of cloth image is carried out three grades of non-lower sampling wavelet transformations, calculate the energy feature of each pixel, these two kinds of features have constituted this width of cloth image texture features collection S jointly.

Step 2 is mapped to a linear measurement space with spectral clustering with feature set S, obtains its mapping ensemblen Y.

1. the image to texture feature extraction uses

The similarity matrix W of approach method calculated characteristics collection S:

(1) adopts any 2 S among the gaussian kernel function calculated characteristics collection S _iWith S _jBetween similarity, that is:

W _ij＝exp(-||s _i-s _j|| ²/2σ ²)

Wherein, σ is a scale parameter.

(2) adopt

Approach method calculates similarity matrix, is that a stochastic sampling n data are as representative point the feature set S of N from size promptly, and the similarity matrix W of formation is:

$W=\left[\begin{array}{cc}A& B\\ B^T& C\end{array}\right]=\left[\begin{array}{cc}A& B\\ B^T& B^T{A}^{-1}B\end{array}\right]$

Wherein, sub-piece A ∈ R ^{N * n}Weight between the representative point of expression stochastic sampling, sub-piece B ∈ R ^{(N-n) * n}The representative point that has comprised stochastic sampling is the weight between the left point in the feature set S, B ^TBe the transposed matrix of B, A ^-1The inverse matrix of expression A, C ∈ R ^{(N-n) * (N-n)}Comprised after the stochastic sampling weight between the left point among the feature set S, when n＜＜during N, sub-piece C can be big especially.

2. calculate Laplce's matrix L of similarity matrix W, adopt following formula to calculate, that is:

L＝D ^-1/2WD ^-1/2

In the formula $D=\left[\begin{array}{cccc}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{A}_{1i}& 0& \·\·\·& 0\\ 0& \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{A}_{2i}& \·\·\·& 0\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ 0& 0& \·\·\·& \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{A}_{\mathrm{ni}}\end{array}\right]$

3. ask preceding c the maximum pairing proper vector of eigenwert of Laplce's matrix L, as the mapping ensemblen Y of characteristic data set in mapping space, c is the image category number;

The eigenwert of matrix, the computing formula of proper vector are: (λ I-L) X=0

In the formula, λ representation feature value, I is a unit matrix, X is a proper vector, and

$\mathrm{\λ}=\left[\begin{array}{c}{\mathrm{\λ}}_1\\ {\mathrm{\λ}}_2\\ \·\·\·\\ {\mathrm{\λ}}_n\end{array}\right]$ $X=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \·\·\·& x_{1n}\\ x_{21}& x_{22}& \·\·\·& x_{2n}\\ x_{11}& \·\·\·& \·\·\·& \·\·\·\\ x_{n1}& x_{n2}& \·\·\·& x_{\mathrm{nn}}\end{array}\right]$

Then mapping ensemblen Y can be expressed as:

$Y=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \·\·\·& x_{1c}\\ x_{21}& x_{22}& \·\·\·& x_{2c}\\ x_{11}& \·\·\·& \·\·\·& \·\·\·\\ x_{n1}& x_{n2}& \·\·\·& x_{\mathrm{nc}}\end{array}\right]$

Step 3, the image division classification number c according to given selects the respective counts destination data at random as initial cluster center from mapping (enum) data collection Y, clone successively, make a variation, select and judge, finds out and the same number of optimum cluster centre of its classification.

1. from mapping ensemblen Y, select c data at random as initial cluster center, and be expressed as:

$B\left(k\right)=\left[\begin{array}{c}{Y}_1\\ Y_2\\ \·\·\·\\ Y_c\end{array}\right]=\left[\begin{array}{cccc}{y}_{11}& y_{12}& \·\·\·& y_{1c}\\ y_{21}& y_{22}& \·\·\·& y_{2c}\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ y_{c1}& y_{c2}& \·\·\·& y_{\mathrm{cc}}\end{array}\right]$

In the formula, k is an iteration step, k=0 during initialization, Y ₁, Y ₂... Y _cThe c that expression is selected from mapping ensemblen Y at random data as initial cluster center;

2. initialized cluster centre B (k) is utilized clone operations operator T _c ^CCarry out clone operations, the cluster centre B ' after obtaining cloning (k) is:

$B^{\′}\left(k\right)=T_c^C\left[B\left(k\right)\right]={\left[\begin{array}{cccc}{Y}_1& Y_1& \·\·\·& Y_1\\ Y_2& Y_2& \·\·\·& Y_2\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ Y_c& Y_c& \·\·\·& Y_c\end{array}\right]}_{(c*p_c)}$

In the formula, (c*p _c) size of representing matrix, p _cBe clone's scale;

3. B ' (k) is carried out cluster centre B after mutation operation obtains making a variation " (k) be:

B″(k)＝B′(k)±rand*p _m

In the formula, rand represents a random number between [0～1], p _mBe the variation probability;

4. " (k) carry out cluster centre B (k+1) after selection operation obtains selecting; promptly calculate data among the mapping ensemblen Y to the cluster centre B " quadratic sum of Euclidean distance (k) to B, as the affinity of this cluster centre, from every class cluster centre, select have maximum affinity cluster centre as B (k+1);

5. repeating step (2)～(4), and according to predefined iteration number iteration step k is judged, if current iteration step greater than predefined iteration number, then stops iteration, otherwise continue to carry out (2).

Step 4 is adjudicated each pixel of image, and the pixel that is about to characteristics of image collection representative is divided in its nearest optimum cluster centre, and each pixel is composed a class mark by the classification of the optimum cluster centre in its place.

1. preliminary ruling: calculate among the mapping ensemblen Y each data point and this data point is adjudicated in the cluster centre with nearest Euclidean distance to the Euclidean distance of the optimum cluster centre that above-mentioned immune cluster obtained;

2. conclusive judgement: owing to the data among the feature set S in data among the Y in mapping space and the former space are one to one, so the court verdict of data is the court verdict of primitive character collection S just accordingly also in the Y set, promptly each pixel of the pairing image of feature set S is endowed corresponding class mark according to the division of Y.

Step 5, the output segmentation result

It is the divided result of presentation video that each pixel is endowed corresponding class mark, the whole output of these class mark pixels is promptly finished the cutting procedure of entire image.

Advantage of the present invention is further specified by the data and the image of following emulation.

1. simulated conditions

(1) four class data: Iris, Sonar, Wine and Segment from the UCI database, have been selected respectively, from the texture image storehouse, selected two width of cloth texture images and two width of cloth SAR images and assessed the performance of different dividing methods, simultaneously with the present invention and traditional based on the k-mean cluster image partition method and compare based on the image partition method of Classical Spectrum cluster.

(2) the immune spectrum clustering algorithm that adopts in the emulation experiment is end condition with the maximum iteration time all, and maximum iteration time was decided to be for 20 generations.

(3) the antibody of cluster centre as the immune clone clustering algorithm, the population scale of this immune cluster clone algorithm is consistent with the clusters number of setting, and clone's scale is 5 times of the antibody scale, i.e. p _c=5, the variation Probability p _m=0.7, use in the experiment based on

Randomly drawing 100 pixels in the approach method finds the solution as representative point.

(4) represent image partition method with K-means based on the k-mean cluster; Represent image partition method with NJW based on the Classical Spectrum cluster; Represent the image partition method based on immune spectrum clustering of the present invention with NEW.

2. emulation content and result

(1) to the emulation experiment of UCI data

This emulation is chosen four data sets respectively and is tested from UCI machine learning database, experimental data is as shown in table 1.

The selected UCI data attribute list of table 1 experiment

With the performance of the present invention of the data test in the table 1, each data is as a feature set, this feature set is shone upon its mapping ensemblen of acquisition earlier, seek the optimum cluster centre of this mapping ensemblen again with immune clone clustering, according to the optimum cluster centre that obtains each data is composed at last and marked with class.Because selected data itself have correct class mark, can the class mark that emulation experiment obtains be compared with real class mark, calculate the average mistake of 10 operations and divide rate, the quality of algorithm is described with average mistake branch rate.Can produce big influence to the result in that the value of simulation process mesoscale parameter σ is different, for Iris,, Sonar, Wine and Segment data, the optimal parameter σ that selects by a large amount of experiments is respectively 1.3,0.44,0.44 and 1.2.

Table 2 has been listed employing K-means, the average mistake branch rate that data set emulation is obtained in three kinds of method his-and-hers watches 1 of NJW and NEW.

The average error rate of table 2UCI data

As seen from Table 2, for the Iris data, the average mistake branch rate of NJW method decreases than K-means, and the average mistake branch rate of immune spectrum clustering method of the present invention has reduced by 8.78% and 8.28% respectively than K-means and NJW method; For the Sonar data, the average mistake of NJW method divides rate slightly higher than K-means, and the average mistake branch rate of immune spectrum clustering method of the present invention has substantial degradation with respect to K-means and NJW method, has reduced by 23.95% and 24.60% respectively; For the Wine data, the average mistake branch rate of NJW method is lower than K-means, and the average mistake branch rate of immune spectrum clustering method of the present invention has reduced by 2.67% and 2.23% respectively than K-means and NJW method; For the Segment data, the average mistake branch rate of NJW method has greatly improved than K-means, and the average mistake branch rate of immune spectrum clustering method of the present invention has reduced by 26.52% than K-means, has reduced by 15.99% than NJW method.In a word, immune spectrum clustering method of the present invention is better than K-means and NJW method to a great extent.

(2) to the emulation experiment of texture image

Get the former figure of synthetic texture image that two width of cloth sizes are 256*256 shown in Fig. 3 (a) and Fig. 4 (a), this two width of cloth texture image comprises 2 classes and 4 class textures respectively, its desirable segmentation result is shown in Fig. 3 (b) and Fig. 4 (b), and the emulation segmentation result is respectively shown in Fig. 3 (c)～Fig. 3 (e) and Fig. 4 (c)～Fig. 4 (e).Wherein, Fig. 3 (c) is the segmentation result of K-means to Fig. 3 (a), Fig. 3 (d) is the segmentation result of NJW method to Fig. 3 (a), Fig. 3 (e) is the segmentation result of immune spectrum clustering method of the present invention to Fig. 3 (a), Fig. 4 (c) is the segmentation result of K-means to Fig. 4 (a), Fig. 4 (d) is the segmentation result of NJW method to Fig. 4 (a), and Fig. 4 (e) is the segmentation result of immune spectrum clustering method of the present invention to Fig. 4 (a).Table 3 has been listed the average mistake branch rate of 10 operations that actual segmentation result cuts apart with respect to ideal, and the optimal parameter σ that selects by a large amount of experiments is respectively 0.2 and 0.15.

The average mistake of table 3 two width of cloth texture images is divided rate

From the segmentation result of the synthetic texture image of two width of cloth as can be seen, the segmentation result of immune spectrum clustering method of the present invention all is better than the result of K-means on edge accuracy and regional consistance, there are serious mistake branch in Fig. 3 (c) and Fig. 4 (c) in some zone especially border, this is because spectral clustering has the ability of the cluster of the non-protruding distribution of identification, is more suitable in the partition problem with various DATA DISTRIBUTION; The segmentation result of immune spectrum clustering method of the present invention also all is better than the NJW method on regional consistance and edge accuracy, there is more mistake branch in zone Fig. 3 (d) compares on the border with Fig. 3 (e) around, and Fig. 4 (d) compares at boundary with Fig. 4 (e) and has more mistake branch.

Average mistake from table 3 is divided rate also as can be seen, and the average mistake branch rate of immune spectrum clustering method of the present invention is starkly lower than K-means and NJW method.For Fig. 3, average mistake the branchs rate of NJW method is lower than K-means, and immune spectrum clustering method of the present invention has reduced by 6.79% and 2.48% respectively than the average branch rate of missing of K-means and NJW method; For Fig. 4, average mistake the branchs rate of NJW method also is lower than K-means, and immune spectrum clustering method of the present invention has reduced by 3.84% and 2.49% respectively than the average branch rate of missing of K-means and NJW method.In a word, no matter from visual angle or the average wrong branch rate of cutting apart with respect to ideal, immune spectrum clustering method of the present invention all is better than K-means and NJW method.

(3) to the emulation experiment of SAR image

Experiment is selected for use as Fig. 5 (a) and Fig. 6 (a) two width of cloth size and is the SAR image of 256*256, and first width of cloth image comprises two class atural objects: vegetation and forest; Second width of cloth image comprises three class atural objects: river, vegetation and forest.Wherein, Fig. 5 (b) is the segmentation result of K-means to Fig. 5 (a), Fig. 5 (c) is the segmentation result of NJW method to Fig. 5 (a), Fig. 5 (d) is the segmentation result of immune spectrum clustering method of the present invention to Fig. 5 (a), Fig. 6 (b) is the segmentation result of K-means to Fig. 6 (a), Fig. 6 (c) is the segmentation result of NJW method to Fig. 6 (a), and Fig. 6 (d) is the segmentation result of immune spectrum clustering method of the present invention to Fig. 6 (a).The optimal parameter σ that selects by a large amount of experiments is respectively 0.2 and 0.3.

As can be seen the mistake of two class atural objects is divided seriously from Fig. 5 (b), as in the zone, upper right side of Fig. 5 (b), vegetation is divided into forest by mistake; The segmentation result of NJW method shown in Fig. 5 (c), is better than the result of K-means, and it all improves a lot on the accuracy on regional consistance and border; Fig. 5 (d) is the segmentation result of immune spectrum clustering method of the present invention, as can be seen, immune spectrum clustering method of the present invention all is better than the segmentation result of K-means to the segmentation result of two class atural objects, compare with the NJW method, Boundary Recognition to forest has improvement slightly, can obtain the correct division in each zone of SAR image.

As can be seen from the figure the mistake of forest is divided seriously, the pipeline on the river is not split simultaneously.Also there is serious mistake branch in vegetation in the lower left of Fig. 6 (b); The segmentation result of NJW method is better than the result of K-means as Fig. 6 (c), yet does not discern the pipeline on the river surface is also correct, and the vegetation of its lower left is divided into the river by mistake simultaneously; Fig. 6 (d) is the segmentation result of immune spectrum clustering of the present invention, as can be seen, Fig. 6 (d) makes moderate progress in the regional consistance in the lower left corner with to the identification of pipeline, therefore the segmentation result of immune spectrum clustering of the present invention is better than K-means, result than NJW method is good slightly, can obtain the correct division in each zone of SAR image.

In sum, the image partition method based on immune spectrum clustering of the present invention is compared with other two kinds of image partition methods Not only reduced intrinsic dimensionality, and improved nicety of grading, segmentation effect is better.

Claims (3)

1. image partition method based on immune spectrum clustering comprises following process:

(1) extracts the textural characteristics of input picture, and each pixel in the image is represented with a proper vector, obtain the characteristics of image collection;

(2) with spectral clustering the characteristics of image collection is mapped to a linear measurement space, obtains its mapping (enum) data collection;

(3) according to given image division classification number, select the respective counts destination data at random as initial cluster center from mapping (enum) data is concentrated, clone successively, make a variation, select and judge, find out and the same number of optimum cluster centre of its classification;

(4) all pixels with the characteristics of image collection are divided in its nearest optimum cluster centre, and each pixel is composed a class mark by the classification of the optimum cluster centre in its place, are the segmentation result of image.

2. image partition method according to claim 1, wherein step (2) is carried out according to following process:

(2a) similarity of computed image feature set, and constitute similarity matrix A, promptly

$A=\begin{array}{}\left[\begin{array}{cccc}{A}_{11}& A_{12}& ...& A_{1n}\\ A_{21}& A_{22}& ...& A_{2n}\\ ...& ...& ...& ...\\ A_{n1}& A_{n2}& ...& A_{\mathrm{nn}}\end{array}\right]\end{array}$

Calculating formula of similarity is

A _ij＝exp(-||s _i-s _j|| ²/2σ ²)

In the formula, A _IjTwo data point s in the presentation video feature set _iWith s _jBetween similarity, σ is a scale parameter, n is the number of characteristics of image collection;

(2b) Laplce's matrix of calculating similarity matrix A: L=D ^-1/2AD ^-1/2

In the formula, $D=\left[\begin{array}{cccc}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{A}_{1i}& 0& ...& 0\\ 0& \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{A}_{2i}& ...& 0\\ ...& ...& ...& ...\\ 0& 0& ...& \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{A}_{\mathrm{ni}}\end{array}\right];$

(2c) ask preceding c of the Laplce's matrix L maximum pairing proper vector of eigenwert, as the mapping ensemblen Y of feature set in mapping space, c is the image category number,

The eigenwert of matrix, the computing formula of proper vector are: (λ I-L) X=0

In the formula, λ representation feature value, I is a unit matrix, X is a proper vector, and

$\mathrm{\λ}=\left[\begin{array}{c}{\mathrm{\λ}}_1\\ {\mathrm{\λ}}_2\\ ...\\ {\mathrm{\λ}}_n\end{array}\right]$ $X=\left[\begin{array}{cccc}{x}_{11}& x_{12}& ...& x_{1n}\\ x_{21}& x_{22}& ...& x_{2n}\\ x_{11}& ...& ...& ...\\ x_{n1}& x_{n2}& ...& x_{\mathrm{nn}}\end{array}\right]$

Then mapping ensemblen Y can be expressed as:

$Y=\begin{array}{c}\left[\begin{array}{cccc}{x}_{11}& x_{12}& ...& x_{1c}\\ x_{21}& x_{22}& ...& x_{2c}\\ x_{11}& ...& ...& ...\\ x_{n1}& x_{n2}& ...& x_{\mathrm{nc}}\end{array}\right]\end{array}.$

3. image partition method according to claim 1, wherein step (3) is carried out according to following process: (3a) select c data at random as initial cluster center from mapping ensemblen Y, and be expressed as:

$B\left(k\right)=\left[\begin{array}{c}{Y}_1\\ Y_2\\ ...\\ Y_c\end{array}\right]=\left[\begin{array}{cccc}{y}_{11}& y_{12}& ...& y_{1c}\\ y_{21}& y_{22}& ...& y_{2c}\\ ...& ...& ...& ...\\ y_{c1}& y_{c2}& ...& y_{\mathrm{cc}}\end{array}\right]$

In the formula, k is an iteration step, k=0 during initialization, Y ₁, Y ₂... Y _cThe c that expression is selected from mapping ensemblen Y at random data as initial cluster center;

(3b) initialized cluster centre B (k) is utilized clone operations operator T _c ^CCarry out clone operations, the cluster centre B ' after obtaining cloning (k) is:

$B^{\′}\left(k\right)=T_c^C\left[B\left(k\right)\right]={\left[\begin{array}{cccc}{Y}_1& Y_1& ...& Y_1\\ Y_2& Y_2& ...& Y_2\\ ...& ...& ...& ...\\ Y_c& Y_c& ...& Y_c\end{array}\right]}_{\left(c^{*}{p}_c\right)}$

In the formula, (c ^*p _c) size of representing matrix, p _cBe clone's scale;

(3c) B ' (k) is carried out cluster centre B after mutation operation obtains making a variation " (k) be:

B″(k)＝B′(k)±rand ^*p _m

In the formula, rand represents a random number between [0～1], p _mBe the variation probability;

" (k) carry out cluster centre B (k+1) after selection operation obtains selecting; promptly calculate data among the mapping ensemblen Y (3d) to B to the cluster centre B " quadratic sum of Euclidean distance (k), as the affinity of this cluster centre, from every class cluster centre, select have maximum affinity cluster centre as B (k+1);

(3e) repeating step (3b)～(3d), and according to predefined iteration number iteration step k is judged, if current iteration step greater than predefined iteration number, then stops iteration, otherwise continue to carry out (3b).

Images