CN108921853B - Image segmentation method based on super-pixel and immune sparse spectral clustering - Google Patents

Image segmentation method based on super-pixel and immune sparse spectral clustering Download PDF

Info

Publication number
CN108921853B
CN108921853B CN201810651506.7A CN201810651506A CN108921853B CN 108921853 B CN108921853 B CN 108921853B CN 201810651506 A CN201810651506 A CN 201810651506A CN 108921853 B CN108921853 B CN 108921853B
Authority
CN
China
Prior art keywords
antibody
pop
super
clustering
antibodies
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810651506.7A
Other languages
Chinese (zh)
Other versions
CN108921853A (en
Inventor
尚荣华
刘爽
焦李成
刘芳
尚凡华
王蓉芳
侯彪
王爽
马文萍
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN201810651506.7A priority Critical patent/CN108921853B/en
Publication of CN108921853A publication Critical patent/CN108921853A/en
Application granted granted Critical
Publication of CN108921853B publication Critical patent/CN108921853B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10004—Still image; Photographic image

Abstract

The invention discloses an image segmentation method based on superpixel and immune sparse spectral clustering, which mainly solves the problems of low segmentation accuracy and poor robustness of the existing image segmentation method. The method comprises the following steps: firstly, performing superpixel division on a texture image, and extracting texture features of the texture image to serve as a feature data set; then, in the spectral clustering process, an immune clone algorithm and sparse representation are combined to search for an optimal similarity matrix of the characteristic data set; and finally, according to the clustering label and the super-pixel marked original image, the texture image is segmented. The method extracts the super pixel blocks of the image as the characteristic data set, and divides the characteristic data set by using an image segmentation method based on immune sparse spectral clustering, so that a more accurate segmentation result is obtained.

Description

Image segmentation method based on super-pixel and immune sparse spectral clustering
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to an image segmentation technology which can be used for target identification.
Background
Common methods for image segmentation include a threshold-based segmentation method, an edge-based segmentation method, a region-based segmentation method, and the like, and with the theoretical development of pattern recognition and artificial intelligence, segmentation methods combining specific theories are also widely applied, such as a clustering-based segmentation technology, an artificial neural network-based segmentation technology, a genetic algorithm-based segmentation technology, and the like. Among the existing clustering methods, spectral clustering is a clustering method with good characteristics of simple implementation, independence from dimension and global optimization. The application of spectral clustering to image segmentation is a popular research direction in the field of image segmentation in recent years. The process of image segmentation is to take each pixel as a data point, and the result of segmentation is to assign a class label to these data points. And the pixels with the same class marks are classified into one class, so that the image is segmented.
The key of spectral clustering is to calculate a similarity matrix of NxN, where N is the number of samples, and the calculation amount is too large for large-scale data.
Juanjuan Luo, Licheng Jiao, and Jose A.Lozano, in the paper "A Sparse Spectral Clustering Framework Multi objective evolution Algorithm" (IEEE Transactions on evolution compatibility 20.3(2016): 418-. The method combines sparse representation with a genetic algorithm NSGA2 of non-dominant ordering to optimize a similarity matrix of spectral clustering. The method improves the generation of a similarity matrix and improves the clustering accuracy of spectral clustering, but the method has poor convergence speed and solution diversity, adopts RC measurement as an index of final solution selection, and has no accurate NC measurement.
The patent of the university of electronic science and technology of xi' an image segmentation method based on immune sparse spectral clustering (patent application No. 200910024374.6, publication No. CN101673398A) discloses an image segmentation method based on immune sparse spectral clustering. The method adopts the combination of an immune clone algorithm and a spectral clustering algorithm to cluster the images, thereby realizing the segmentation of the images. Although the method does not need prior knowledge and has low calculation complexity, the method still has the defect of low segmentation precision.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides an image segmentation method based on superpixel and immune sparse spectral clustering to extract more concise and effective features and improve the segmentation precision.
The basic idea of the invention is as follows: performing superpixel division on the texture image, and extracting texture features of the texture image to obtain a clustering data set; searching an optimal clustering data set similarity matrix by using an immune clone algorithm and sparse representation in the spectral clustering process; and (4) segmenting the texture image by combining the clustering label with the super-pixel label original image.
According to the above thought, the implementation steps of the invention include the following:
(1) inputting an image to be segmented with the size of 256 x 256;
(2) dividing the image to be segmented into super pixels by using a simple linear iterative clustering SLIC method;
(3) extracting the characteristics of the superpixels by using a gray level co-occurrence matrix and gabor wavelet transform;
(4) normalizing all superpixel feature data to be between [0,1] to remove the influence of magnitude between data to obtain a normalized superpixel feature data set I;
(5) obtaining a non-dominant antibody set P by using an immune clone selection method for the normalized super-pixel characteristic data set I;
(6) sequentially taking each non-dominant antibody of the non-dominant antibody set P as a similarity matrix of a spectral clustering process, and performing spectral clustering on the super-pixel characteristic data set I to obtain a series of clustering results related to the super-pixel characteristic data set I;
(7) calculating the NC measurement of each clustering result, and selecting the clustering result corresponding to the smallest NC measurement as the final clustering result of the super-pixel characteristic data set I;
(8) and according to the clustering result of the super-pixel characteristic data set I, marking the category of each pixel point as the category of the corresponding super-pixel point, and taking the category as the clustering result of the image to be segmented to finish the image segmentation.
Compared with the prior art, the invention has the following advantages:
firstly, in the clustering process of image segmentation, the invention adopts superpixel processing to obtain a representative superpixel block, overcomes the defect of large number of samples when the original image characteristics are used for spectral clustering, and reduces the calculated amount;
secondly, in the clustering process of image segmentation, the immune clone algorithm is combined with sparse spectral clustering, so that widely distributed and various candidate solutions can be obtained, and the segmentation precision of the image is improved;
thirdly, in the clustering process of image segmentation, the method adopts NC measurement to select the final solution, overcomes the defect that the similarity in the image is not considered in RC measurement, and improves the accuracy of image segmentation.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention;
FIG. 2 is a comparison graph of the segmentation of a class 2 texture image using the two methods of the present invention and the prior art texture image segmentation;
FIG. 3 is a comparison graph of the segmentation of a class 3 texture image using the present invention and two existing texture image segmentation methods.
Detailed Description
The present invention is described in further detail below with reference to the attached drawing figures.
Step 1, inputting an image to be segmented.
And selecting a texture image with the size of A-256 from the database as an input image to be segmented.
And 2, dividing the image to be segmented into superpixels.
The existing methods for dividing superpixels include a graph theory-based method, an entropy rate-based method, a gradient descent-based method Meanshift, a simple linear iterative clustering method SLIC, a level set method Turbopixels based on geometric flow and the like. In the embodiment, a simple linear iterative clustering method SLIC is used for dividing an image to be segmented into n super-pixel points, and the implementation steps are as follows:
(2a) uniformly selecting n pixel points as initial seed points in an image to be segmented, distributing a label for each initial seed point, wherein each seed point comprises A/n pixel points, and the distance between the seed points is
Figure BDA0001704958530000031
Figure BDA0001704958530000032
(2b) Reselecting the seed points in the 2S-2S neighborhood of the seed points, namely calculating gradient values of all pixel points in the neighborhood, moving the seed points to the place with the minimum gradient in the neighborhood, and avoiding the position of the seed points falling on the image boundary from interfering the clustering process;
(2c) in the 2S-2S neighborhood of each seed point, calculating the distance D' between the pixel point in the neighborhood and the seed point,
Figure BDA0001704958530000033
wherein d iscIndicating color distance
Figure BDA0001704958530000034
lj、ajAnd bjRespectively representing color information of seed points,/i、aiAnd biRespectively representing color information of pixel points, dsRepresenting spatial distances
Figure BDA0001704958530000035
xjAnd yjSpatial information, x, representing seed points, respectivelyiAnd yiRespectively representing the space information of the pixel points, wherein m represents a balance parameter used for measuring the proportion of the color value and the space information in similarity measurement, and setting the label of each pixel point as the label of the seed point closest to the pixel point;
(2d) continuously iterating the processes (2b) and (2c) until convergence, obtaining n seed points, using the seed points as superpixels, and obtaining a set V ═ V { (V) of the superpixels1,v2,…,vi,…,vn}，viRepresenting the ith superpixel point.
And 3, extracting a characteristic data set of the super pixels.
The method for extracting the characteristics of the texture image comprises a gray level co-occurrence matrix, wavelet transformation, non-downsampling wavelet transformation and the like, the method adopted by the embodiment is gray level co-occurrence matrix and gabor wavelet transformation, and the method is realized as follows:
(3a) obtaining 8 characteristics of all pixel points of the image to be segmented through the gray level co-occurrence matrix:
(3a1) vectorizing an image to be segmented into 16 gray levels;
(3a2) generating a gray level co-occurrence matrix p (i, j) in four directions of which the distance d is equal to 1, the direction is theta 0 degrees, theta 45 degrees, theta 90 degrees and theta 135 degrees in a window of 5 x 5 for each pixel point of an image to be segmented, wherein i and j are index values of the gray level co-occurrence matrix in the transverse direction and the longitudinal direction respectively;
(3a3) calculating four statistics of contrast, energy, entropy and correlation of the gray level co-occurrence matrix:
contrast ratio:
Figure BDA0001704958530000041
energy:
Figure BDA0001704958530000042
entropy:
Figure BDA0001704958530000043
correlation:
Figure BDA0001704958530000044
wherein p' (i, j) is normalized gray level co-occurrence matrix, μx、σxRespectively represent the mean and variance, μ, of p' (i, j) along the x-axisy、σyThe mean and variance of p' (i, j) along the y-axis direction are respectively represented;
(3a4) respectively calculating the contrast, energy, entropy and the mean and variance of correlation of the four gray level co-occurrence matrixes to obtain 8 feature vectors of all pixel points of the image to be segmented;
(3b) obtaining 2 characteristics of all pixel points of the image to be segmented through gabor wavelet transform:
transforming the image to be segmented into a frequency domain space by using two different gabor filter groups, extracting texture features of the image to be segmented in two different scales and different directions, wherein the two different gabor filter groups are G respectively1And G2Expressed as follows:
Figure BDA0001704958530000045
G2＝G·cos[2·π·f·(x·cosθ+y·sinθ)]
wherein G (x, y) represents a Gabor function,
Figure BDA0001704958530000051
sx、syrepresenting the variance along the x-axis and y-axis, respectively, f being the sine functionThe frequency of the number theta represents the direction of the gabor filter, and 2 feature vectors of all pixel points of the image to be segmented are obtained together;
(3c) combining 8 features of all pixel points acquired by the gray level co-occurrence matrix and 2 features acquired by gabor wavelet transformation, acquiring 10 features of each pixel point, averaging texture feature data of a plurality of pixel points included in each super pixel point, and taking the average as a feature data value of the super pixel point to obtain a super pixel feature data set H with the size of n x 10, wherein n represents the number of the super pixel points, and 10 represents the number of the features.
And 4, normalizing the super-pixel characteristic data set H to be between [0 and 1] to remove the influence of magnitude between data to obtain a normalized super-pixel characteristic data set I.
(4a) Selecting any element in the super-pixel characteristic data set H;
(4b) calculating the difference between the element and the minimum element in the column of the element;
(4c) calculating the difference value between the maximum element in the row of the element and the minimum element in the row of the element;
(4d) dividing the difference obtained in the step (4b) by the difference obtained in the step (4c) to obtain a normalization result of the taken element;
(4e) and (4) judging whether the elements in the super-pixel characteristic data set H are completely taken, if so, combining the normalization results of all the elements into a normalized super-pixel characteristic data set I, and if not, returning to the step (4 a).
Step 5, non-dominant antibody populations are obtained using an immune cloning algorithm.
(5a) Initializing antibody population:
the maximum number of iterations T is set to 20, the antibody population size N is set to 30, and an initial antibody population pop (T) (X) is randomly generated1， X2，…，Xl，…，XN) Where t represents the number of iterations, initialized to 0, XlThe first antibody is represented by 1, …, N, N represents the size of the antibody population,
Figure BDA0001704958530000052
xijrepresents the ith superThe similarity between the pixel and the jth super pixel, i is 1, …, n, j is 1, …, n, n represents the number of super pixels,
Figure BDA0001704958530000053
rand is a random number between 0 and 1;
(5b) calculate the objective function for each antibody:
Figure BDA0001704958530000054
wherein f is1(Xl) Represents antibody XlError term objective function of f2(Xl) Represents antibody XlI represents the characteristic data of the superpixel, | · | | luminance2 2Represents the square of 2 norm, | ·| non-woven phosphor0Represents a norm of 0;
(5c) selection operation
Selection of two antibodies X from the antibody population POPlAnd XmJudging the first antibody XlAnd a second antibody XmWhether or not (f) is satisfied1(Xl)≤f1(Xm)&f2(Xl)<f2(Xm) Or (f)1(Xl)<f1(Xm)&f2(Xl)≤f2(Xm))：
If so, antibody XlIs a non-dominant antibody, which is capable of dominating antibody Xm；
Otherwise, antibody XlNot the dominant antibody, it cannot dominate antibody Xm；
Wherein f is1(Xl) And f2(Xl) Respectively represent the first antibody XlOf the error term objective function and of the sparse term objective function, f1(Xm) And f2(Xm) Respectively represent the second antibody XmThe values of the error term objective function and the sparse term objective function of (a);
selecting all antibodies from the group of POPsNon-dominant antibodies, constituting a non-dominant antibody population POP1；
(5d) Cloning operations
For non-dominant antibody population POP1The antibody population generated by 4 cloning operations is shown as:
POP2＝{POP1，POP1，POP1，POP1}。
(5e) crossover operation
(5e1) Setting two random numbers from 0 to 1, namely a first random number alpha, a second random number beta and a third random number eta;
(5e2) comparing the first random number with a set threshold of 0.5:
if α is>0.5, from the cloned antibody population POP2Two antibodies X are randomly selectedr1And Xr2Uniform crossover was performed, generating new antibodies: xr3，Xr3Element (1) of
Figure BDA0001704958530000061
If alpha is less than or equal to 0.5, the cloned antibody population POP2Two antibodies X are randomly selectedr1And Xr2Arithmetic crossover was performed to generate new antibodies: xr4＝ηXr1+(1-η)Xr2；
(5e3) Repeating the operation (5e2) N times to generate N new antibodies to form the crossed POP3。
(5f) Mutation operation
(5f1) Selecting crossed antibody group POP3Any one of the antibodies XlLet the antibody XlSimilarity x of ith and jth super pixel points in (1)ijThe variation criterion is as follows:
Figure BDA0001704958530000062
wherein rand and gamma are random numbers from 0 to 1, and h is the neighbor order of the jth superpixel point relative to the ith superpixel point;
(5f2) for crossed antibody group POP3The N antibodies are subjected to mutation operation in sequence according to a mutation rule to generate N new antibodies to form a mutated antibody group POP4。
(5g) Judging whether the iteration times reach the maximum iteration times T, if so, carrying out the POP on the antibody group4Step (6) is performed as the final antibody population P, otherwise, step (5b) is returned for the next iteration.
And 6, performing spectral clustering on the non-dominant antibody group P obtained by the immune clone algorithm.
(6a) For antibody X in non-dominant antibody population P l1, …, N, calculation degree matrix
Figure BDA0001704958530000071
Wherein
Figure BDA0001704958530000072
Wherein xijExpressing the similarity between the ith super pixel point and the jth super pixel point, wherein n expresses the number of the super pixel points;
(6b) calculating the Laplace matrix L ═ D-Xl；
(6c) Performing eigenvalue decomposition on the Laplace matrix L according to an eigenvalue decomposition formula Lu ═ λ u to obtain a vector λ consisting of n eigenvalues, λ ═ λ1，λ2，…，λi，…，λn]And n eigenvectors, u ═ u [ u ]1， u2，…，ui，…，un]Wherein the characteristic value λiCorresponding to the feature vector ui；
(6d) Sorting the eigenvalues of the vector lambda from small to large to obtain a vector lambda 'formed by n sorted eigenvalues, wherein lambda' is [ lambda ═ lambda1’，λ2’，…，λn’]Selecting the first h minimum eigenvalues to form a vector lambda', lambda ═ lambda1’，λ2’，…，λh’]The corresponding eigenvectors form a matrix u', u ═ u [ u ]1’，u2’，…，uh’]Normalized to the matrix uObtaining a matrix u' through quantization;
(6e) dividing u' by using a kmeans method to obtain a clustering result V of the super-pixel characteristic data set I, wherein V is (V)1，…，Vi，…，Vn)，ViAnd representing the i-th class of super pixel point set in the clustering result.
And 7, calculating the NC measurement of each clustering result, and selecting the clustering result corresponding to the smallest NC measurement as the clustering result of the super-pixel characteristic data set I.
(7a) Calculating the NC measure of each clustering result by the following formula:
Figure BDA0001704958530000073
where c represents the number of classes of the cluster,
Figure BDA0001704958530000074
the weight between the classes is represented by,
Figure BDA0001704958530000075
Vhrepresenting the h-th class of super-pixel point set in the clustering result,
Figure BDA0001704958530000076
representing a set of superpixels, w, in the clustering result that do not belong to the h-th classijIndicating a standard similarity between the ith superpixel point and the jth superpixel point,
Figure BDA0001704958530000081
Figure BDA0001704958530000082
represents antibody XlSimilarity between the ith and jth superpixel points, where l is 1, …, N denotes the number of clustering results generated by non-dominant antibodies, vol (V)h) Represents a subset VhThe weight of (a) is determined,
Figure BDA0001704958530000083
dithe ith element in the representation matrix;
(7b) and after the NC measurement of all the clustering results is calculated, selecting the clustering result corresponding to the smallest NC measurement as the clustering result of the super-pixel data set I.
And 8, marking the category of each pixel point as the category of the corresponding super pixel point according to the clustering result of the super pixel characteristic data set I, and taking the category as the clustering result of the image to be segmented to finish the image segmentation.
The effects of the present invention can be further illustrated by the following simulations:
1. simulation environment:
the simulation of the invention is carried out in the environment that the computer is configured with core i53.30GHZ, the memory 8G, the WINDOWS 7 system and the computer software are configured with Matlab R2014 b.
2. Simulation content:
the method of the invention, the K mean method and the fuzzy C mean method in the prior art are applied to carry out simulation experiments for segmenting two types and three types of texture images respectively, and the performance of the image segmentation methods is evaluated from the aspects of segmentation precision, edge preservation, detail retention and the like.
Simulation 1, the selected texture image which can be divided into two types is divided by the invention, the result is shown in figure 2, wherein
FIG. 2(a) is a texture image that can be divided into two classes;
FIG. 2(b) is a graph of a simulation segmentation result obtained by segmenting FIG. 2(a) by using the method of the present invention;
FIG. 2(c) is a graph of simulated segmentation results obtained by segmenting FIG. 2(a) using K-means values;
FIG. 2(d) is a simulation segmentation result graph obtained by segmenting the graph of FIG. 2(a) by using a fuzzy C-means clustering algorithm;
FIG. 2(e) is a partition template of a texture map that can be partitioned into two classes.
Simulation 2, the method is adopted to segment a selected texture image which can be segmented into three types, and the result is shown in fig. 3, wherein:
FIG. 3(a) is a texture map divided into three categories;
FIG. 3(b) is a graph of a simulation segmentation result obtained by segmenting FIG. 3(a) using the method of the present invention;
FIG. 3(c) is a graph of simulated segmentation results obtained by segmenting FIG. 3(a) using K-means values;
FIG. 3(d) is a simulation segmentation result graph obtained by segmenting the graph of FIG. 3(a) by using a fuzzy C-means clustering algorithm;
fig. 3(e) is a partition template of a texture image map that can be partitioned into three classes.
The three methods are used for calculating and counting the segmentation accuracy results of the two texture images, and the results are shown in table 1.
TABLE 1 segmentation accuracy of the three methods for these two texture images
The invention K mean value Fuzzy C mean
Two types of texture images 0.9804 0.9629 0.9485
Three types of texture images 0.8964 0.8026 0.7655
3. And (3) simulation result analysis:
as can be seen from fig. 2 and table 1, fig. 2(b) is better than the segmentation results in fig. 2(C) and fig. 2(d), and fig. 2(C) and fig. 2(d) have more outliers than fig. 2(b), but the edge of fig. 2(C) is not as sharp as fig. 2(b), which shows that the present invention has stronger robustness and can obtain more accurate image segmentation results compared with the existing K-means and fuzzy C-means clustering algorithms.
As can be seen from fig. 3 and table 1, fig. 3(b) is better than the segmentation results in fig. 3(C) and fig. 3(d), and fig. 3(C) and fig. 3(d) have more outliers than fig. 3(b), but the edge of fig. 3(C) is not as sharp as fig. 3(b), which shows that the present invention has stronger robustness on segmentation of three types of texture images and can obtain more accurate image segmentation results compared with the existing K-means and fuzzy C-means clustering algorithms.
In conclusion, compared with the conventional K-mean and fuzzy C-mean clustering algorithm, the method disclosed by the invention has the advantage that the accuracy is greatly improved.

Claims (6)

1. An image segmentation method based on superpixel and immune sparse spectral clustering comprises the following steps:
(1) inputting an image to be segmented with the size of 256 x 256;
(2) dividing the image to be segmented into super pixels by using a simple linear iterative clustering SLIC method;
(3) extracting the characteristics of the superpixels by using a gray level co-occurrence matrix and gabor wavelet transform;
(4) normalizing all superpixel feature data to be between [0,1] to remove the influence of magnitude between data to obtain a normalized superpixel feature data set I;
(5) obtaining a non-dominant antibody set P by using an immune clone selection method for the normalized super-pixel characteristic data set I; the method comprises the following steps:
(5a) initializing antibody population:
randomly generating an initial population of antibodies POP (t) ═ X1，X2，…，Xl，…，XN) Where t represents the number of iterations, initialised to0，XlThe first antibody is represented by 1, …, N, N represents the size of the antibody population,
Figure FDA0003343661750000011
xijthe similarity between the ith super pixel point and the jth super pixel point is shown, i is 1, …, n, j is 1, …, n, n is the number of the super pixel points,
Figure FDA0003343661750000012
rand is a random number between 0 and 1;
(5b) calculate the objective function for each antibody:
Figure FDA0003343661750000013
wherein f is1(Xl) Represents antibody XlError term objective function of f2(Xl) Represents antibody XlI represents the characteristic data of the superpixel,
Figure FDA0003343661750000014
represents the square of 2 norm, | ·| non-woven phosphor0Represents a norm of 0;
(5c) selection operation
Selection of two antibodies X from the antibody population POPlAnd XmJudgment of antibody XlAnd antibody XmWhether or not f is satisfied1(Xl)≤f1(Xm)&f2(Xl)<f2(Xm) Or satisfy f1(Xl)<f1(Xm)&f2(Xl)≤f2(Xm)：
If so, antibody XlIs a non-dominant antibody, which is capable of dominating antibody Xm；
Otherwise, antibody XlNot the dominant antibody, it cannot dominate antibody Xm；
Wherein f is1(Xl) And f2(Xl) Respectively represent the first antibody XlOf the error term objective function and of the sparse term objective function, f1(Xm) And f2(Xm) Respectively represent the second antibody XmThe values of the error term objective function and the sparse term objective function of (a);
selecting all non-dominant antibodies from the antibody group POP to form the non-dominant antibody group POP1；
(5d) Immunological manipulation
For non-dominant antibody population POP1Cloning to generate cloned antibody POP2；
For antibody group POP2Performing cross operation to generate cross antibody group POP3；
For antibody group POP3Performing mutation operation to generate mutated antibody group POP4；
(5e) Judging whether the iteration times reach the maximum iteration times T, if so, carrying out the POP on the antibody group4Step (6) is executed as the final antibody group P, otherwise, the step (5b) is returned to for the next iteration;
(6) sequentially taking each non-dominant antibody of the non-dominant antibody set P as a similarity matrix of a spectral clustering process, and performing spectral clustering on the super-pixel characteristic data set I to obtain a series of clustering results related to the super-pixel characteristic data set I;
(7) calculating the NC measurement of each clustering result, and selecting the clustering result corresponding to the smallest NC measurement as the final clustering result of the super-pixel characteristic data set I;
(8) and according to the final clustering result of the super-pixel characteristic data set I, marking the category of each pixel point as the category of the corresponding super-pixel point, and taking the category as the clustering result of the image to be segmented to finish the image segmentation.
2. The method of claim 1, wherein the antibody population POP after cloning in step (5d)2Is to the non-dominant antibody population POP1The antibody population generated by 4 cloning operations is shown as:
POP2＝{POP1，POP1，POP1，POP1}。
3. the method of claim 1, wherein the cross-linked population of antibodies POP is generated in step (5d)3It is implemented as follows:
firstly, setting two random numbers from 0 to 1, namely a first random number alpha, a second random number beta and a third random number eta;
and secondly, comparing the first random number with a set threshold value of 0.5:
if α is>0.5, from the cloned antibody population POP2Two antibodies X are randomly selectedr1And Xr2Uniform crossover was performed, generating new antibodies: xr3The antibody Xr3The similarity between the ith and jth superpixel points in (1) can be calculated as
Figure FDA0003343661750000021
If alpha is less than or equal to 0.5, the cloned antibody population POP2Two antibodies X are randomly selectedr1And Xr2Arithmetic crossover was performed to generate new antibodies: xr4＝ηXr1+(1-η)Xr2；
Thirdly, repeating the second step for N times to generate N new antibodies to form a crossed antibody group POP3。
4. The method of claim 1, wherein the variant population of antibodies POP is generated in step (5d)4It is implemented as follows:
firstly, selecting crossed antibody group POP3Any one of the antibodies XlLet the antibody XlSimilarity x of ith and jth super pixel points in (1)ijThe variation criterion is as follows:
Figure FDA0003343661750000031
wherein rand and gamma are random numbers from 0 to 1, and h is the neighbor order of the jth superpixel point relative to the ith superpixel point;
then, the crossed antibody group POP is treated3The N antibodies are subjected to mutation operation in sequence according to a mutation rule to generate N new antibodies to form a mutated antibody group POP4。
5. The method of claim 1, wherein the spectral clustering of the superpixel feature dataset I in step (6) is performed by:
(6a) for antibody X in non-dominant antibody population Pl1, …, N, calculation degree matrix
Figure FDA0003343661750000032
Wherein
Figure FDA0003343661750000033
Wherein xijExpressing the similarity between the ith super pixel point and the jth super pixel point, wherein n expresses the number of the super pixel points;
(6b) calculating the Laplace matrix L ═ D-Xl；
(6c) And decomposing the eigenvalue of the Laplace matrix L according to a formula Lu- λ u to obtain an eigenvalue λ and an eigenvector u-u1，u2，…，un]；
(6d) Selecting the eigenvectors corresponding to the first h smallest eigenvalues to form u' ═ u1’，u2’，…，uh’]Normalizing u' to obtain a normalized feature vector u ";
(6e) and (3) dividing u' by using a kmeans method to obtain a clustering result of the super-pixel characteristic data set I:
V＝(V1，…，Vi，…，Vc)，
wherein ViAnd c, representing the super pixel point set belonging to the ith class in the clustering result, and representing the number of the clustered classes.
6. The method of claim 1, wherein step (7) calculates the NC measure for each cluster result by the following formula:
Figure FDA0003343661750000041
where c represents the number of classes of the cluster,
Figure FDA0003343661750000042
the weight between the classes is represented by,
Figure FDA0003343661750000043
Vhrepresenting the h-th class of super-pixel point set in the clustering result,
Figure FDA0003343661750000044
representing a set of superpixels, w, in the clustering result that do not belong to the h-th classijIndicating a standard similarity between the ith superpixel point and the jth superpixel point,
Figure FDA0003343661750000045
Figure FDA0003343661750000046
represents antibody XlSimilarity between the ith and jth superpixel points, where l is 1, …, N denotes the number of clustering results generated by non-dominant antibodies, vol (V)h) Represents a subset VhThe weight of (a) is determined,
Figure FDA0003343661750000047
dithe ith element in the representation matrix.
CN201810651506.7A 2018-06-22 2018-06-22 Image segmentation method based on super-pixel and immune sparse spectral clustering Active CN108921853B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810651506.7A CN108921853B (en) 2018-06-22 2018-06-22 Image segmentation method based on super-pixel and immune sparse spectral clustering

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810651506.7A CN108921853B (en) 2018-06-22 2018-06-22 Image segmentation method based on super-pixel and immune sparse spectral clustering

Publications (2)

Publication Number Publication Date
CN108921853A CN108921853A (en) 2018-11-30
CN108921853B true CN108921853B (en) 2022-03-04

Family

ID=64422288

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810651506.7A Active CN108921853B (en) 2018-06-22 2018-06-22 Image segmentation method based on super-pixel and immune sparse spectral clustering

Country Status (1)

Country Link
CN (1) CN108921853B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109858529B (en) * 2019-01-11 2022-11-01 广东工业大学 Scalable image clustering method
CN112396619B (en) * 2019-08-15 2022-05-10 电子科技大学 Small particle segmentation method based on semantic segmentation and internally complex composition
CN112784894B (en) * 2021-01-18 2022-11-15 西南石油大学 Automatic labeling method for rock slice microscopic image

Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101299243A (en) * 2008-06-27 2008-11-05 西安电子科技大学 Method of image segmentation based on immune spectrum clustering
CN101667292A (en) * 2009-09-25 2010-03-10 西安电子科技大学 SAR image segmentation system and segmentation method based on immune clone and projection pursuit
CN101673398A (en) * 2009-10-16 2010-03-17 西安电子科技大学 Method for splitting images based on clustering of immunity sparse spectrums
CN101866489A (en) * 2010-06-30 2010-10-20 西安电子科技大学 Image dividing method based on immune multi-object clustering
CN102184524A (en) * 2011-04-13 2011-09-14 西安电子科技大学 Neighbourhood learning meme image segmentation method based on standard cut
CN102254326A (en) * 2011-07-22 2011-11-23 西安电子科技大学 Image segmentation method by using nucleus transmission
CN102867307A (en) * 2012-09-10 2013-01-09 西安电子科技大学 SAR image segmentation method based on feature vector integration spectral clustering
CN102968796A (en) * 2012-11-30 2013-03-13 西安电子科技大学 SAR (Synthetic Aperture Radar) image segmentation method based on sampling learning
CN103310441A (en) * 2013-05-16 2013-09-18 西安电子科技大学 Texture image segmentation method based on immunity cloning and multitarget optimizing
CN103366365A (en) * 2013-06-18 2013-10-23 西安电子科技大学 SAR image varying detecting method based on artificial immunity multi-target clustering
CN104156943A (en) * 2014-07-14 2014-11-19 西安电子科技大学 Multi-target fuzzy cluster image variance detecting method based on non-control-neighborhood immune algorithm
CN104217436A (en) * 2014-09-16 2014-12-17 西安电子科技大学 SAR image segmentation method based on multiple feature united sparse graph
CN106780639A (en) * 2017-01-20 2017-05-31 中国海洋大学 Hash coding method based on the sparse insertion of significant characteristics and extreme learning machine
CN107784284A (en) * 2017-10-24 2018-03-09 哈尔滨工业大学深圳研究生院 Face identification method and system

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8428363B2 (en) * 2011-04-29 2013-04-23 Mitsubishi Electric Research Laboratories, Inc. Method for segmenting images using superpixels and entropy rate clustering
JP6639113B2 (en) * 2015-06-05 2020-02-05 キヤノン株式会社 Image recognition device, image recognition method, and program

Patent Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101299243A (en) * 2008-06-27 2008-11-05 西安电子科技大学 Method of image segmentation based on immune spectrum clustering
CN101667292A (en) * 2009-09-25 2010-03-10 西安电子科技大学 SAR image segmentation system and segmentation method based on immune clone and projection pursuit
CN101673398A (en) * 2009-10-16 2010-03-17 西安电子科技大学 Method for splitting images based on clustering of immunity sparse spectrums
CN101866489A (en) * 2010-06-30 2010-10-20 西安电子科技大学 Image dividing method based on immune multi-object clustering
CN102184524A (en) * 2011-04-13 2011-09-14 西安电子科技大学 Neighbourhood learning meme image segmentation method based on standard cut
CN102254326A (en) * 2011-07-22 2011-11-23 西安电子科技大学 Image segmentation method by using nucleus transmission
CN102867307A (en) * 2012-09-10 2013-01-09 西安电子科技大学 SAR image segmentation method based on feature vector integration spectral clustering
CN102968796A (en) * 2012-11-30 2013-03-13 西安电子科技大学 SAR (Synthetic Aperture Radar) image segmentation method based on sampling learning
CN103310441A (en) * 2013-05-16 2013-09-18 西安电子科技大学 Texture image segmentation method based on immunity cloning and multitarget optimizing
CN103366365A (en) * 2013-06-18 2013-10-23 西安电子科技大学 SAR image varying detecting method based on artificial immunity multi-target clustering
CN104156943A (en) * 2014-07-14 2014-11-19 西安电子科技大学 Multi-target fuzzy cluster image variance detecting method based on non-control-neighborhood immune algorithm
CN104217436A (en) * 2014-09-16 2014-12-17 西安电子科技大学 SAR image segmentation method based on multiple feature united sparse graph
CN106780639A (en) * 2017-01-20 2017-05-31 中国海洋大学 Hash coding method based on the sparse insertion of significant characteristics and extreme learning machine
CN107784284A (en) * 2017-10-24 2018-03-09 哈尔滨工业大学深圳研究生院 Face identification method and system

Non-Patent Citations (7)

* Cited by examiner, † Cited by third party
Title
A Sparse Spectral Clustering Framework via Multi-Objective Evolutionary Algorithm;Juanjuan Luo等;《IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION》;20150903;第20卷(第3期);第1-17页 *
Improved multi-objective clustering algorithm using particle swarm optimization;Congcong Gong等;《PLOS ONE》;20171205;第12卷(第12期);第1-19页 *
Multiobjective evolutionary algorithms: A comparative case study and he strength Pareto approach;EcKart Zitzler等;《IEEE Transactions on Evolutionary Computation》;19991130;第3卷(第4期);第257-271页 *
图像分割的改进稀疏子空间聚类方法;李小平等;《系统工程与电子技术》;20151031;第37卷(第10期);第2418-2424页 *
基于人工免疫系统的多目标优化与SAR图像分割;杨咚咚;《中国优秀博硕士学位论文全文数据库(博士) 信息科技辑》;20130415(第04期);第I140-7页 *
融合SLIC的DCUT改进图像分割算法;邹小林;《新疆大学学报(自然科学版)》;20170228;第34卷(第1期);第78-83、 *
遗传算法优化归一化划分准则的图像分割;翟艳鹏等;《计算机工程与应用》;20101231;第46卷(第33期);第148-150、157页 *

Also Published As

Publication number Publication date
CN108921853A (en) 2018-11-30

Similar Documents

Publication Publication Date Title
Yang et al. A feature-reduction multi-view k-means clustering algorithm
Dong et al. Color clustering and learning for image segmentation based on neural networks
CN105427296B (en) A kind of thyroid gland focus image-recognizing method based on ultrasonoscopy low rank analysis
CN108537102B (en) High-resolution SAR image classification method based on sparse features and conditional random field
CN111080684B (en) Point cloud registration method for point neighborhood scale difference description
CN109543723B (en) Robust image clustering method
CN107203785A (en) Multipath Gaussian kernel Fuzzy c-Means Clustering Algorithm
CN108921853B (en) Image segmentation method based on super-pixel and immune sparse spectral clustering
CN104732545B (en) The texture image segmenting method with quick spectral clustering is propagated with reference to sparse neighbour
Khan et al. Color image segmentation: a novel spatial fuzzy genetic algorithm
CN110796667A (en) Color image segmentation method based on improved wavelet clustering
CN110188763B (en) Image significance detection method based on improved graph model
CN110866439B (en) Hyperspectral image joint classification method based on multi-feature learning and super-pixel kernel sparse representation
CN109840518B (en) Visual tracking method combining classification and domain adaptation
EP1353295B1 (en) Non-linear data mapping and dimensionality reduction system
CN114840925A (en) Method for registering vehicle body part measurement data to integral CAD model
CN110619311A (en) Data classification method based on EEMD-ICA-SVM
CN111126467B (en) Remote sensing image space spectrum clustering method based on multi-target sine and cosine algorithm
CN106951873B (en) Remote sensing image target identification method
CN112509017A (en) Remote sensing image change detection method based on learnable difference algorithm
CN111539966A (en) Colorimetric sensor array image segmentation method based on fuzzy c-means clustering
CN111539444A (en) Gaussian mixture model method for modified mode recognition and statistical modeling
CN107492101B (en) Multi-modal nasopharyngeal tumor segmentation algorithm based on self-adaptive constructed optimal graph
Ganebnykh et al. Metric classifier using multilevel network of templates
CN105844299A (en) Image classification method based on bag of words

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant