CN104732545B - The texture image segmenting method with quick spectral clustering is propagated with reference to sparse neighbour - Google Patents
The texture image segmenting method with quick spectral clustering is propagated with reference to sparse neighbour Download PDFInfo
- Publication number
- CN104732545B CN104732545B CN201510155156.1A CN201510155156A CN104732545B CN 104732545 B CN104732545 B CN 104732545B CN 201510155156 A CN201510155156 A CN 201510155156A CN 104732545 B CN104732545 B CN 104732545B
- Authority
- CN
- China
- Prior art keywords
- data point
- similarity matrix
- sparse
- point
- gray scale
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 46
- 230000003595 spectral effect Effects 0.000 title claims abstract description 10
- 230000000644 propagated effect Effects 0.000 title description 3
- 239000011159 matrix material Substances 0.000 claims abstract description 108
- 230000011218 segmentation Effects 0.000 claims abstract description 47
- 230000015572 biosynthetic process Effects 0.000 claims abstract description 39
- 239000013598 vector Substances 0.000 claims description 14
- 238000013480 data collection Methods 0.000 claims description 6
- 239000012141 concentrate Substances 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- ODKSFYDXXFIFQN-UHFFFAOYSA-M argininate Chemical compound [O-]C(=O)C(N)CCCNC(N)=N ODKSFYDXXFIFQN-UHFFFAOYSA-M 0.000 claims 1
- 238000005516 engineering process Methods 0.000 abstract description 7
- 238000000926 separation method Methods 0.000 abstract description 2
- 238000003709 image segmentation Methods 0.000 description 8
- 230000000694 effects Effects 0.000 description 5
- 238000000605 extraction Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000000513 principal component analysis Methods 0.000 description 2
- 238000001228 spectrum Methods 0.000 description 2
- 239000005557 antagonist Substances 0.000 description 1
- 238000013475 authorization Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 230000007115 recruitment Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The present invention discloses the texture image segmenting method of a kind of sparse neighbour's propagation of combination and quick spectral clustering, mainly solves the problems, such as that the accurate low and computation complexity of segmentation of existing texture image segmenting method is high.Its method and step is：(1) it is input into an image to be split；(2) arrange parameter；(3) generation gray scale is to probability square formation；(4) number at statistical number strong point；(5) sparse similarity matrix is built；(6) data are chosen to represent a little；(7) data point is clustered；(8) image to be split is marked；(9) image after output segmentation.The present invention can preferably keep the region consistency of texture image relative to existing some Study Of Segmentation Of Textured Images technologies, and the acquisition accuracy of separation is high, the texture image of segmentation good stability, and computation complexity is low.
Description
Technical field
The invention belongs to technical field of image processing, the sparse neighbour of combination of technical field of image segmentation is further related to
Propagate the texture image segmenting method with quick spectral clustering.The present invention can be used for the segmentation of various texture images, to reach identification
With the purpose of analysis target.
Background technology
Image segmentation is one of basic problem of image procossing, is the basis for realizing carrying out image target identification.Wherein,
Study Of Segmentation Of Textured Images are an important branch of Digital Image Processing research, are the bases of numerous graphical analyses and machine vision applications
Plinth.
In image partition method, the Study Of Segmentation Of Textured Images of feature based generally by feature extraction and image segmentation the two
Step is successively performed.The purpose of Study Of Segmentation Of Textured Images is according to spies such as gray scale, color, texture, space, the geometries of image
Levy and divide an image into several mutually disjoint regions, it is desirable to which the inside in each region has consistent texture, without same district
Texture between domain is different.In the texture image segmentation of feature based, each pixel is used one by feature extraction
Individual textural characteristics are represented, then feature set to obtaining is divided to realize Study Of Segmentation Of Textured Images.In this regard, clustering ensemble technology
There is provided a kind of approach of Study Of Segmentation Of Textured Images scheme, the precision of image segmentation can be effectively improved.FCM Algorithms (FCM
Algorithm) it is a kind of most easy and effective, most widely used clustering method and is frequently used for Study Of Segmentation Of Textured Images.FCM algorithm bases
In fuzzy theory, objective world can more accurately be described, and algorithm is simple, fast convergence rate.But, the algorithm is to initial
Cluster centre is sensitive and is easily trapped into local optimum, limits the accuracy of cluster result.In addition, spectral clustering be also it is a kind of most
Conventional clustering method, although this clustering effect is pretty good, and can solve the problems, such as non-convex data set.But, this
Algorithm needs to construct adjacent degree matrix when splitting for texture maps, and needs to calculate the spy of its corresponding Laplacian Matrix
Levy vector.The time complexity of this two step is respectively O (n2) and O (n3).This computation complexity high causes that it can only solve small rule
The clustering problem of mould data set.When in face of large-scale dataset, not only need to take considerable time, and can take in a large amount of
Deposit.So, when Study Of Segmentation Of Textured Images are processed with the method, the feature of substantial amounts can behave clustering algorithm becomes non-
It is often time-consuming.
The patent " texture image segmenting method based on immune clone multiple-objection optimization " of Xian Electronics Science and Technology University's application
Disclose a kind of excellent based on immune clone multiple target in (number of patent application 201310182014.5, publication number CN103310441A)
The texture image segmenting method of change.The step of realizing of the method is：Step 1, is input into texture image to be split, and extract its feature
Matrix G；Step 2, produces initial antibodies group V (t) and is initially set；Step 3, according to eigenmatrix G and antibody population V (t),
Calculate cluster object function f1With class object function f2；Step 4, antagonist group V (t) carries out immune clone operation, is cloned
Antibody population Vc (t) afterwards；Step 5, nonuniform meshes operation is carried out to antibody population Vc (t) after clone, obtains non-uniform change
Antibody population Vm (t) after different；Step 6, population recruitment operation is carried out to antibody population Vm (t) after non-uniform variation, is updated
Antibody population Vm (t+1) afterwards；Step 7, according to antibody population Vm (t+1) and eigenmatrix G after renewal, in calculating texture image
The classification of each pixel；Step 8, the texture image after output segmentation.Although the method can preferably keep texture image
Region consistency, improves segmentation precision, segmentation result is more met the vision of people, but the deficiency being still present is：Due to exempting from
Epidemic disease clone algorithm belongs to evolution algorithm, when Study Of Segmentation Of Textured Images are carried out with immune clone algorithm, the stability of segmentation result
Difference, it is impossible to obtain effective image segmentation result.In addition, the evolutionary generation of population clone, variation and population easily increases iteration
The time of computing and local extremum is absorbed in, limits the accuracy of cluster result.
Patented technology " a kind of dividing method of texture image " (number of patent application that Shenzhen Graduate School of Tsinghua University possesses
201210259652.8, publication number CN102819840A, Authorization Notice No. 102819840B) in disclose a kind of texture image
Dividing method.The patented technology is comprised the concrete steps that：It is input into image to be split；Extract the local various features information table of image
The characteristic for levying different texture region in image obtains characteristic image, by the means such as principal component analysis reduction data volume again by equal
Value drift algorithm is clustered to characteristic vector, so as to complete relatively reliable Study Of Segmentation Of Textured Images.Although the patented technology is relative
The image segmentation reliability obtained in prior art is high, but the deficiency being still present is：The patented technology is extraction image
Local feature information characterize the overall situation, the randomness of extraction effect is big, is easily lost important information, cause segmentation effect poor.
In addition, local optimum is easily trapped into by the method for principal component analysis means reduction data volume, and marginal information is easily lost,
Limit the precision of segmentation result.
The content of the invention
It is an object of the invention to overcome the shortcomings of above-mentioned prior art, it is proposed that propagated with reference to sparse neighbour and quick spectrum
The texture image segmenting method of cluster, faster more accurately extracts the feature of texture image so as to carry out Study Of Segmentation Of Textured Images, carries
Study Of Segmentation Of Textured Images effect high.
The present invention realizes that the thinking of above-mentioned purpose is：First, texture image to be split is read in, using generation gray scale to general
The single gray feature of image is converted to provincial characteristics by the method for rate square formation；Secondly, recurrence tune is carried out using to K mean cluster
The gray scale of acquisition is divided into K different region to probability square formation by method, counts the number at each number of regions strong point, and count
Other Euclidean distances between putting in each data point and its region are calculated, so as to build a sparse similarity matrix；So
Afterwards, sparse similarity matrix is acted on using neighbour's propagation algorithm, obtains a representative point composition and represent matrix；Finally, with quick spectrum
Clustering algorithm is clustered to data point, exports the segmentation result of image to be split.
It is to achieve these goals, of the invention that to implement step as follows：
(1) it is input into an image to be split：
It is 256 × 256 texture images to be split of pixel to be input into a width size；
(2) arrange parameter：
The gray scale numerical series of the texture image to be split being input into is set to 16, maximum iteration is set to 60；
(3) generation gray scale is to probability square formation：
(3a) sets up the plane coordinate system of texture image to be split with the central point of texture image to be split as origin；
(3b) any ash chosen at 2 points, read corresponding to 2 points for choosing from texture image plane coordinate system to be split
2 points of corresponding gray values are constituted a gray scale pair by angle value size；
(3c) moves in parallel at 2 points that are used for obtaining gray scale pair in step (3b) on whole coordinate plane, often moves one
One gray scale pair of secondary acquisition, 16 are obtained on the dynamic face of translation altogether2Plant gray scale pair；
(3d) in whole texture image coordinate plane to be split 2562Individual gray scale centering, statistics 162Plant gray scale centering each
Plant the number of gray scale pair；
The number of each gray scale on texture image coordinate plane to be split is arranged in a square formation by (3e), by the square formation
Normalization, obtains gray scale to probability square formation；
(4) number at statistical number strong point:
(4a) uses K mean cluster method, and gray scale is divided into K different region, K >=2 to probability square formation；
Each region is divided into smaller region, until all areas by (4b) in each region recursive call K mean cluster method
The data amount check in domain is less than or equal to untill K；
(4c) is counted and is preserved the number at each number of regions strong point；
(5) sparse similarity matrix is built：
(5a) calculates other Euclidean distances between putting in each data point and its region；
(5b) chooses the most short point of Euclidean distance of each data point；
(5c) using all of data point as sparse similarity matrix first row, the Euclidean distance of each data point is most short
Point as sparse similarity matrix secondary series, the Euclidean distance between two columns strong points is the 3rd of sparse similarity matrix
Row；
(6) data are chosen to represent a little：
Using neighbour's propagation algorithm, the representative point of element in sparse similarity matrix is calculated；
(7) data point is clustered：
The point that represents of element in sparse similarity matrix is arranged in a size as p × p matrix V, p represents sparse by (7a)
The number of the representative point of element, 0 in similarity matrix<p≤2000；
(7b) according to the following formula, calculates the degree of membership of cluster data collection：
Wherein, uijRepresent that cluster data concentrates j-th data membership in i-th degree of membership of class, i=1,2 ...,
65536, j=1,2 ..., p, p represent the number of the representative point of element in sparse similarity matrix, 0<P≤2000, exp is represented and referred to
Number operation, dijRepresent gray scale the i-th line number strong point in probability square formation to the Euclidean distance between jth line number strong point in matrix V, δ tables
Show Gauss nuclear parameter, δ=100, ∑ represents sum operation；
The angle value that is subordinate to of the cluster data collection of acquisition is arranged in a size for 65536 × p subordinated-degree matrix by (7c), and 0<
p≤2000；
(7d) calculates the p dimensional feature vectors A of subordinated-degree matrix, 0<p≤2000；
(7e) is clustered using K mean cluster method to characteristic vector A, obtains the cluster label of characteristic vector A；
(8) image to be split is marked：
Using the cluster label of characteristic vector as each pixel of image category label, in obtaining texture image to be split
The category label of each pixel；
(9) image after output segmentation：
Texture image to be split is split according to category label, the image after being split simultaneously exports segmentation result.
Compared with prior art, the present invention has advantages below：
First, because the present invention is using method of the gray scale to probability square formation is generated, by the single gray feature of texture image
Provincial characteristics is converted to, not directly with the gray value of image as characteristic data set, prior art is overcome to texture image
Computationally intensive shortcoming when being split so that present invention reduces computation complexity, improve processing speed.
Second, the method due to the present invention using recursive call is carried out to K mean cluster, the gray scale that will be obtained is to probability side
Battle array is divided into the different regions of K, overcomes that prior art accuracy of separation when Study Of Segmentation Of Textured Images are carried out is low, image easy to lose
The shortcoming of marginal information so that the present invention improves the accuracy to Study Of Segmentation Of Textured Images.
3rd, because the present invention acts on sparse similarity matrix using neighbour's propagation algorithm, obtain and represent point composition generation
Table matrix overcomes Spectral Clustering calculating high when image segmentation is applied in the prior art as the input of spectral clustering
Complexity this shortcoming, while it is sensitive to initial cluster center when Study Of Segmentation Of Textured Images are carried out and easily to overcome prior art
It is absorbed in the shortcoming of local optimum so that the present invention improves the accuracy to Study Of Segmentation Of Textured Images and to obtain stability more preferable
Segmentation result.
Brief description of the drawings
Fig. 1 is flow chart of the invention；
Fig. 2 is analogous diagram of the invention, and wherein Fig. 2 (a) is the artificial synthesized two kinds of texture maps of different texture that contain, figure
2 (b) is the result split to artificial synthesized texture maps using standard FCM texture image segmenting methods of the prior art
Figure, Fig. 2 (c) is the result figure split to artificial synthesized texture maps using the method for the present invention.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.
It is of the invention to comprise the following steps that referring to the drawings 1.
Step 1, one width size of input is 256 × 256 texture images to be split of pixel.
Step 2, arrange parameter.
The gray scale numerical series of the texture image to be split being input into is set to 16, maximum iteration is set to 60.
Step 3, generation gray scale is to probability square formation.
Central point with texture image to be split sets up the plane coordinate system of texture image to be split as origin.
Any gray value chosen at 2 points, read corresponding to 2 points for choosing from texture image plane coordinate system to be split
2 points of corresponding gray values are constituted a gray scale pair by size.
2 points that gray scale pair will be used for obtaining are moved in parallel on whole coordinate plane, often mobile once to obtain a gray scale
It is right, 16 are obtained altogether on the dynamic face of translation2Plant gray scale pair.
The 256 of whole texture image coordinate plane to be split2Individual gray scale centering, statistics 162Each is grey to plant gray scale centering
Spend to number.
The number of each gray scale on texture image coordinate plane to be split is arranged in a square formation, by the square formation normalizing
Change, to probability square formation, the gray scale of acquisition is 256 × 256 to the size of probability square formation to generation gray scale.
Step 4, the number at statistical number strong point.
Using K mean cluster method, gray scale is divided into K different region, K >=2 to probability square formation.
The step of K mean cluster method carries out region division to gray scale to probability square formation be：From the gray scale pair of image to be split
K element is randomly choosed in probability square formation as initial cluster center value, each cluster centre is each divided into a class, K >=
2, the gray scale of image to be split is calculated to all elements in probability square formation to the K distance of cluster centre value, compare gray scale to general
Each element is to the K distance of cluster centre value in rate square formation, by the corresponding cluster centre value classification mark of each element minimum value
Number corresponding element is assigned, obtain category label of the gray scale to each element in probability square formation, calculate gray scale in probability square formation
Average value per dvielement, obtains new cluster centre value, and whether relatively newer cluster centre value is identical with former cluster centre value,
If new cluster centre value is different with original cluster centre value, continue iteration, by gray scale to all elements in probability square formation according to
Distance with new cluster centre value re-starts category division, until reaching maximum iteration, cluster result is exported, if newly
Cluster centre value is identical with former cluster centre value, then export cluster result.
Gray scale is calculated to each element in probability square formation to K cluster centre value using Euclidean distance in K mean cluster method
Distance, Euclidean distance calculates according to the following formula：
D (x, y)=| | x-y | |.
Wherein, d (x, y) represents Euclidean distance, and x and y represents gray scale to two elements in probability square formation respectively.
In each region recursive call K mean cluster method, each region is divided into smaller region, until all regions
Data amount check counts and preserves the number at each number of regions strong point less than or equal to untill K.
Step 5, builds sparse similarity matrix.
Other Euclidean distances between putting in each data point and its region are calculated, the Euclidean of each data point is chosen
The most short point of distance.
Using all of data point as sparse similarity matrix first row, the most short point of the Euclidean distance of each data point
Used as the secondary series of sparse similarity matrix, the Euclidean distance between two columns strong points is the 3rd row of sparse similarity matrix.
The present invention is to the specific embodiment for building sparse similarity matrix：Assuming that the neighbour of data point 1 is 2,3,4, number
The neighbour at strong point 2 is 3,5,6, then sparse similarity matrix can be expressed as：
Wherein, d (x, y) represents Euclidean distance, and x and y represents gray scale to two elements in probability square formation respectively.
Step 6, chooses data and represents a little.
Using neighbour's propagation algorithm, the representative point of element in sparse similarity matrix is calculated.
Neighbour's propagation algorithm is comprised the following steps that：
1st step, at the beginning of the Attraction Degree and degree of membership between element in sparse similarity matrix are initialized as into 0, iterations t
Beginning turns to 1；
2nd step, according to the following formula, Attraction Degrees of the data point j to data point i in the sparse similarity matrix of calculating：
Wherein, r (i, j) represents in sparse similarity matrix that to the Attraction Degree of data point i, i and j are represented data point j respectively
I-th and j-th label of element in sparse similarity matrix, i=1,2 ..., 65536, j=1,2 ..., 65536, s (i, j)
The similarity size between data point i and data point j in sparse similarity matrix is represented, max is represented and taken maxima operation, k tables
Show k-th label of element in sparse similarity matrix, k=1,2 ..., 65536, t represent current iterations, t≤60, a
(i, k) represents the degree of membership of data point i in sparse similarity matrix to data point k, and s (i, k) is represented in sparse similarity matrix
Similarity size between data point i and data point k；
3rd step, according to the following formula, degree of membership of the data point i to data point j in the sparse similarity matrix of calculating：
Wherein, a (i, j) represents in sparse similarity matrix that to the degree of membership of data point j, i and j are represented data point i respectively
I-th and j-th label of element in sparse similarity matrix, i=1,2 ..., 65536, j=1,2 ..., 65536, min tables
Show and take minimum Value Operations, r (j, j) represent sparse similarity matrix in data point j to the Attraction Degree of itself, ∑ represent summation behaviour
Make, t represents current iterations, t≤60, and max is represented and taken maxima operation, r (k, j) represents number in sparse similarity matrix
Strong point k represents k-th label of element in sparse similarity matrix, k=1,2 ..., 65536 to the Attraction Degree of data point j, k；
4th step, according to the following formula, Attraction Degrees of the data point j to data point i in the sparse similarity matrix of renewal：
R (i, j)=λ × r'(i, j)+(1- λ) × r (i, j).
Wherein, after R (i, j) is represented and updated in sparse similarity matrix data point j to the Attraction Degree of data point i, i and j points
Do not represent i-th and j-th label of element in sparse similarity matrix, i=1,2 ..., 65536, j=1,2 ..., 65536,
λ represents convergence coefficient, λ ∈ (0,1), r'(i, before j) representing current iteration in sparse similarity matrix data point j to data point
The Attraction Degree of i, r (i, j) be in the sparse similarity matrix that the 2nd step is obtained data point j to the Attraction Degree of data point i；
5th step, according to the following formula, degree of membership of the data point i to data point j in the sparse similarity matrix of renewal：
A (i, j)=λ × a'(i, j)+(1- λ) × a (i, j).
Wherein, after A (i, j) is represented and updated in sparse similarity matrix data point i to the degree of membership of data point j, i and j points
Do not represent i-th and j-th label of element in sparse similarity matrix, i=1,2 ..., 65536, j=1,2 ..., 65536,
λ represents convergence coefficient, λ ∈ (0,1), a'(i, before j) representing current iteration in sparse similarity matrix data point i to data point
The degree of membership of j, a (i, j) be in the sparse similarity matrix that the 3rd step is obtained data point i to the degree of membership of data point j；
6th step, judges whether current iterations reaches the maximum iteration 60 set by initialization, if it is,
The 7th step is performed, otherwise, the 2nd step is performed；
7th step, according to the following formula, calculates the representative point of element in sparse similarity matrix：
i*=argmax ((A (i, j)+R (i, j)).
Wherein, i*The representative point of data point i in sparse similarity matrix is represented, i is represented i-th in sparse similarity matrix
The label of element, i=1,2 ..., 65536, max are represented and take maxima operation, and arg (i) is represented and taken in sparse similarity matrix
The most operation of value element correspondence label, A (i, j) represent iteration update after in sparse similarity matrix data point i to data point j's
Degree of membership, j represents j-th label of element in sparse similarity matrix, and j=1,2 ..., 65536, R (i, j) represents iteration more
Attraction Degrees of the data point j to data point i in sparse similarity matrix after new.
Neighbour's propagation algorithm need not specify clusters number, but be automatically determined by the size of bias, in this hair
In bright embodiment, bias is set as the median of sparse similarity matrix.
Step 7, clusters to data point.
The point that represents of acquisition is arranged in a size as p × p matrix V, p represents the number for representing point, 0<p≤2000.
According to the following formula, calculate cluster data collection is subordinate to angle value：
Wherein, uijRepresent that cluster data concentrates j-th data membership in i-th degree of membership of class, i=1,2 ...,
65536, j=1,2 ..., p, p represent the number of the representative point of element in sparse similarity matrix, 0<P≤2000, exp is represented and referred to
Number operation, dijRepresent gray scale the i-th line number strong point in probability square formation to the Euclidean distance between jth line number strong point in matrix V, δ tables
Show Gauss nuclear parameter, δ=100, ∑ represents sum operation.
The angle value that is subordinate to of the cluster data collection of acquisition is arranged in a size for 65536 × p subordinated-degree matrix, 0<p≤
2000, calculate the p dimensional feature vectors A of subordinated-degree matrix, 0<P≤2000, are carried out using K mean cluster method to characteristic vector A
Cluster, obtains the cluster label of characteristic vector A.
Step 8, marks image to be split.
Using the cluster label of characteristic vector as each pixel of image category label, in obtaining texture image to be split
The category label of each pixel.
Step 9, marks image to be split.
Texture image to be split is split according to category label, the image after being split simultaneously exports segmentation result.
Effect of the invention can be further illustrated by following emulation：
1. simulated conditions：
L-G simulation test of the invention is to be configured to AMD FX (tm) -6300@3.50GHz, 16.0GB in computer hardware
The hardware environment and computer software of RAM are configured under the software environment of Matlab R2010a what is carried out.
2. emulation content：
Fig. 2 is analogous diagram of the invention, wherein, Fig. 2 (a) is the width people of containing the different texture of two classes of any selection
Work synthesizes texture maps.
Emulation 1, using the standard FCM texture image segmenting methods of the prior art line different to the selected class of a width two
Reason image graph 2 (a) is split, as a result as shown in Fig. 2 (b).Emulation 2, it is different to the selected class of a width two using the present invention
Texture image Fig. 2 (a) split, as a result as shown in Fig. 2 (c).
3. analysis of simulation result：
Fig. 2 (b) is different to the selected class of a width two using the standard FCM texture image segmenting methods of prior art
Texture image split the segmentation result figure for obtaining, and from Fig. 2 (b), segmenting edge is relatively rough, segmentation result it is white
There are many miscellaneous points in color region, there is no good holding area uniformity.Fig. 2 (c) is to selected using the inventive method
The different texture image of the class of one width two split the segmentation result figure for obtaining, from Fig. 2 (c), segmentation knot of the invention
Fruit is preferably maintained to the region consistency after Study Of Segmentation Of Textured Images, and segmenting edge is more regular, and segmentation result more meets people
Visual experience.
Statistics is different to the selected class of a width two using the present invention and standard FCM texture image segmenting methods respectively
Texture image split in the segmentation result for obtaining the correct number of cut-point and normalizes, and obtains using both approaches pair
The segmentation accuracy that the different texture image of the selected class of a width two is split, as a result as shown in table 1.
Table 1 is using the present invention texture maps different to the selected class of a width two with standard FCM texture image segmenting methods
As the segmentation accuracy list split
| Simulation type | The present invention | Standard FCM methods |
| Segmentation accuracy | 0.9714 | 0.9061 |
As known from Table 1, the segmentation split using the present invention texture image different to the selected class of a width two is just
True rate is 0.9714, it is clear that the line different to the selected class of a width two higher than the FCM texture image segmenting methods of the standard of use
The segmentation accuracy 0.9061 that reason image is split, illustrates that the present invention improves the segmentation precision of Study Of Segmentation Of Textured Images.
Claims (3)
1. a kind of sparse neighbour of combination propagates the texture image segmenting method with quick spectral clustering, comprises the following steps：
(1) it is input into an image to be split：
It is 256 × 256 texture images to be split of pixel to be input into a width size；
(2) arrange parameter：
The gray scale numerical series of the texture image to be split being input into is set to 16, maximum iteration is set to 60；
(3) generation gray scale is to probability square formation：
(3a) sets up the plane coordinate system of texture image to be split with the central point of texture image to be split as origin；
(3b) any gray value chosen at 2 points, read corresponding to 2 points for choosing from texture image plane coordinate system to be split
2 points of corresponding gray values are constituted a gray scale pair by size；
(3c) moves in parallel at 2 points that are used for obtaining gray scale pair in step (3b) on whole coordinate plane, and often movement is once obtained
A gray scale pair is obtained, 16 are obtained altogether on the dynamic face of translation2Plant gray scale pair；
(3d) in whole texture image coordinate plane to be split 2562Individual gray scale centering, statistics 162Each is grey to plant gray scale centering
Spend to number；
The number of each gray scale on texture image coordinate plane to be split is arranged in a square formation by (3e), by the square formation normalizing
Change, obtain gray scale to probability square formation；
(4) number at statistical number strong point:
(4a) uses K mean cluster method, and gray scale is divided into K different region, K >=2 to probability square formation；
Each region is divided into smaller region, until all regions by (4b) in each region recursive call K mean cluster method
Data amount check is less than or equal to untill K；
(4c) is counted and is preserved the number at each number of regions strong point；
(5) sparse similarity matrix is built：
(5a) calculates other Euclidean distances between putting in each data point and its region；
(5b) chooses the most short point of Euclidean distance of each data point；
(5c) using all of data point as sparse similarity matrix first row, the most short point of the Euclidean distance of each data point
Used as the secondary series of sparse similarity matrix, the Euclidean distance between two columns strong points is the 3rd row of sparse similarity matrix；
(6) data are chosen to represent a little：
Using neighbour's propagation algorithm, the representative point of element in sparse similarity matrix is calculated；
(7) data point is clustered：
The point that represents of element in sparse similarity matrix is arranged in a size as p × p matrix V, p represents sparse similar by (7a)
The number of the representative point of element, 0 ＜ p≤2000 in degree matrix；
(7b) according to the following formula, calculates the degree of membership of cluster data collection：
Wherein, uijRepresent that cluster data concentrates j-th data membership in i-th degree of membership of class, i=1,2 ..., 65536, j=
1,2 ..., p, p represent the number of the representative point of element in sparse similarity matrix, and 0 ＜ p≤2000, exp represents index operation,
dijGray scale is represented the i-th line number strong point in probability square formation to the Euclidean distance between jth line number strong point in matrix V, δ represents Gaussian kernel
Parameter, δ=100, Σ represents sum operation；
The angle value that is subordinate to of the cluster data collection of acquisition is arranged in a size for 65536 × p subordinated-degree matrix by (7c), and 0 ＜ p≤
2000；
(7d) calculates the p dimensional feature vectors A of subordinated-degree matrix, 0 ＜ p≤2000；
(7e) is clustered using K mean cluster method to characteristic vector A, obtains the cluster label of characteristic vector A；
(8) image to be split is marked：
Using the cluster label of characteristic vector as the category label of each pixel of image, each in texture image to be split is obtained
The category label of pixel；
(9) image after output segmentation：
Texture image to be split is split according to category label, the image after being split simultaneously exports segmentation result.
2. the sparse neighbour of combination according to claim 1 propagates the texture image segmenting method with quick spectral clustering, and it is special
Levy and be, K mean cluster method comprises the following steps that described in step (4a), step (4b)：
1st step, the gray scale from image to be split, will to randomly choosing K element in probability square formation as initial cluster center value
Each cluster centre is each divided into a class, K >=2；
2nd step, calculates the gray scale of image to be split to all elements in probability square formation to the K distance of cluster centre value；
3rd step, compares gray scale to each element in probability square formation to the K distance of cluster centre value, by each element minimum value
Corresponding cluster centre value category label assigns corresponding element, obtains classification mark of the gray scale to each element in probability square formation
Number；
4th step, calculates gray scale to the average value in probability square formation per dvielement, obtains new cluster centre value；
Whether the 5th step, relatively newer cluster centre value is identical with former cluster centre value, if new cluster centre value and former cluster centre
Value is different, then continue iteration, and gray scale is entered again to all elements in probability square formation according to the distance with new cluster centre value
Row category division, until reaching maximum iteration, exports cluster result, if new cluster centre value and former cluster centre value phase
Together, then cluster result is exported.
3. the sparse neighbour of combination according to claim 1 propagates the texture image segmenting method with quick spectral clustering, and it is special
Levy and be, obtained using neighbour's propagation algorithm described in step (6) and represent comprising the following steps that for point：
1st step, 0, iterations t initialization are initialized as by the Attraction Degree and degree of membership between element in sparse similarity matrix
It is 1；
2nd step, according to the following formula, Attraction Degrees of the data point j to data point i in the sparse similarity matrix of calculating：
Wherein, r (i, j) represents the Attraction Degree of data point j in sparse similarity matrix to data point i, and i and j represents sparse respectively
I-th and j-th label of element in similarity matrix, i=1,2 ..., 65536, j=1,2 ..., 65536, s (i, j) is represented
Similarity size in sparse similarity matrix between data point i and data point j, max is represented and is taken maxima operation, and k represents dilute
Dredge k-th label of element in similarity matrix, k=1,2 ..., 65536, t represent current iterations, t≤60, a (i,
K) degree of membership of data point i in sparse similarity matrix to data point k is represented, s (i, k) represents data in sparse similarity matrix
Similarity size between point i and data point k；
3rd step, according to the following formula, degree of membership of the data point i to data point j in the sparse similarity matrix of calculating：
Wherein, a (i, j) represents the degree of membership of data point i in sparse similarity matrix to data point j, and i and j represents sparse respectively
I-th and j-th label of element in similarity matrix, i=1,2 ..., 65536, j=1,2 ..., 65536, min represent and take
Minimum Value Operations, r (j, j) represent sparse similarity matrix in data point j to the Attraction Degree of itself, Σ represents sum operation, t tables
Show current iterations, t≤60, max is represented and taken maxima operation, and r (k, j) represents data point k in sparse similarity matrix
To the Attraction Degree of data point j, k represents k-th label of element in sparse similarity matrix, k=1,2 ..., 65536；
4th step, according to the following formula, Attraction Degrees of the data point j to data point i in the sparse similarity matrix of renewal：
R (i, j)=λ × r'(i, j)+(1- λ) × r (i, j)
Wherein, after R (i, j) is represented and updated in sparse similarity matrix data point j to the Attraction Degree of data point i, i and j difference tables
Show i-th and j-th label of element in sparse similarity matrix, i=1,2 ..., 65536, j=1,2 ..., 65536, λ tables
Show convergence coefficient, λ ∈ (0,1), r'(i, before j) representing current iteration in sparse similarity matrix data point j to data point i's
Attraction Degree, r (i, j) be in the sparse similarity matrix that the 2nd step is obtained data point j to the Attraction Degree of data point i；
5th step, according to the following formula, degree of membership of the data point i to data point j in the sparse similarity matrix of renewal：
A (i, j)=λ × a'(i, j)+(1- λ) × a (i, j)
Wherein, after A (i, j) is represented and updated in sparse similarity matrix data point i to the degree of membership of data point j, i and j difference tables
Show i-th and j-th label of element in sparse similarity matrix, i=1,2 ..., 65536, j=1,2 ..., 65536, λ tables
Show convergence coefficient, λ ∈ (0,1), a'(i, before j) representing current iteration in sparse similarity matrix data point i to data point j's
Degree of membership, a (i, j) be in the sparse similarity matrix that the 3rd step is obtained data point i to the degree of membership of data point j；
6th step, judges whether current iterations reaches the maximum iteration 60 set by initialization, if it is, performing
7th step, otherwise, performs the 2nd step；
7th step, according to the following formula, calculates the representative point of element in sparse similarity matrix：
i*=argmax ((A (i, j)+R (i, j))
Wherein, i*The representative point of data point i in sparse similarity matrix is represented, i represents i-th element in sparse similarity matrix
Label, i=1,2 ..., 65536, max are represented and take maxima operation, and arg () is represented and taken most in sparse similarity matrix
The operation of value element correspondence label, data point i returns to data point j in sparse similarity matrix after A (i, j) expression iteration renewals
Category degree, j represents j-th label of element in sparse similarity matrix, j=1, and 2 ..., 65536, R (i, j) represents that iteration updates
Attraction Degrees of the data point j to data point i in sparse similarity matrix afterwards.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510155156.1A CN104732545B (en) | 2015-04-02 | 2015-04-02 | The texture image segmenting method with quick spectral clustering is propagated with reference to sparse neighbour |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510155156.1A CN104732545B (en) | 2015-04-02 | 2015-04-02 | The texture image segmenting method with quick spectral clustering is propagated with reference to sparse neighbour |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104732545A CN104732545A (en) | 2015-06-24 |
| CN104732545B true CN104732545B (en) | 2017-06-13 |
Family
ID=53456413
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510155156.1A Expired - Fee Related CN104732545B (en) | 2015-04-02 | 2015-04-02 | The texture image segmenting method with quick spectral clustering is propagated with reference to sparse neighbour |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104732545B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106485686A (en) * | 2015-08-17 | 2017-03-08 | 西安电子科技大学 | One kind is based on gravitational spectral clustering image segmentation algorithm |
| CN105760900B (en) * | 2016-04-08 | 2019-06-18 | 西安电子科技大学 | Hyperspectral image classification method based on neighbour's propagation clustering and sparse Multiple Kernel Learning |
| CN107578063B (en) * | 2017-08-21 | 2019-11-26 | 西安电子科技大学 | Image Spectral Clustering based on fast selecting landmark point |
| CN107657228B (en) * | 2017-09-25 | 2020-08-04 | 中国传媒大学 | Video scene similarity analysis method and system, and video encoding and decoding method and system |
| WO2019154444A2 (en) * | 2018-05-31 | 2019-08-15 | 上海快仓智能科技有限公司 | Mapping method, image acquisition and processing system, and positioning method |
| CN211668521U (en) * | 2018-05-31 | 2020-10-13 | 上海快仓智能科技有限公司 | Automatic guide vehicle for image acquisition and processing system |
| CN109858529B (en) * | 2019-01-11 | 2022-11-01 | 广东工业大学 | Scalable image clustering method |
| CN110427956B (en) * | 2019-04-18 | 2021-01-15 | 中国林业科学研究院资源信息研究所 | LiDAR point cloud data single tree extraction method based on spectral clustering algorithm |
| CN112750127B (en) * | 2021-02-04 | 2022-08-26 | 深圳市泽峰光电科技有限公司 | Image processing method for log end face measurement |
| CN114359023B (en) * | 2022-01-10 | 2022-11-18 | 成都智元汇信息技术股份有限公司 | Method, equipment and system for dispatching picture shunt to center based on complexity |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102096825A (en) * | 2011-03-23 | 2011-06-15 | 西安电子科技大学 | Graph-based semi-supervised high-spectral remote sensing image classification method |
| CN103678949A (en) * | 2014-01-09 | 2014-03-26 | 江南大学 | Tracking measurement set partitioning method for multiple extended targets based on density analysis and spectrum clustering |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8798349B2 (en) * | 2011-07-05 | 2014-08-05 | The Chinese University Of Hong Kong | Systems and methods for detecting arterial input function (AIF) |
-
2015
- 2015-04-02 CN CN201510155156.1A patent/CN104732545B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102096825A (en) * | 2011-03-23 | 2011-06-15 | 西安电子科技大学 | Graph-based semi-supervised high-spectral remote sensing image classification method |
| CN103678949A (en) * | 2014-01-09 | 2014-03-26 | 江南大学 | Tracking measurement set partitioning method for multiple extended targets based on density analysis and spectrum clustering |
Non-Patent Citations (4)
| Title |
|---|
| Constrainted spectral clustering through affinity propagation;Zhengdong Lu et al.;《2008 IEEE Conference on Computer Vision and Pattern Recognition》;20081231;全文 * |
| Finding image exemplars using fast sparse affinity propagation;Yangqing Jia et al.;《MM" 08 Proceedings of the 16th ACM international conference onmultimedia 》;20081031;全文 * |
| 基于快速谱聚类的图像分割算法;李纯 等;《应用科技》;20120430(第2期);全文 * |
| 基于稀疏表示的近邻传播聚类算法;胡晨晓 等;《西南大学学报（自然科学版）》;20140531;第36卷(第5期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104732545A (en) | 2015-06-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104732545B (en) | The texture image segmenting method with quick spectral clustering is propagated with reference to sparse neighbour | |
| CN107358293B (en) | Neural network training method and device | |
| CN112184752A (en) | Video target tracking method based on pyramid convolution | |
| CN106650744B (en) | The image object of local shape migration guidance is divided into segmentation method | |
| CN103593855B (en) | The image partition method of cluster is estimated based on particle group optimizing and space length | |
| CN108959379B (en) | Garment image retrieval method based on visual salient region and hand-drawn sketch | |
| CN108416347A (en) | Well-marked target detection algorithm based on boundary priori and iteration optimization | |
| de Arruda et al. | A complex networks approach for data clustering | |
| CN110443809A (en) | Structure sensitive property color images super-pixel method with boundary constraint | |
| Narayana et al. | Instantaneous approach for evaluating the initial centers in the agricultural databases using K-means clustering algorithm | |
| CN108846404A (en) | A kind of image significance detection method and device based on the sequence of related constraint figure | |
| CN108921853B (en) | Image segmentation method based on super-pixel and immune sparse spectral clustering | |
| CN106340004A (en) | Fuzzy clustering preprocessing cloud system-based parallel cloud drift wind inversion method | |
| He et al. | A Method of Identifying Thunderstorm Clouds in Satellite Cloud Image Based on Clustering. | |
| CN107301643A (en) | Well-marked target detection method based on robust rarefaction representation Yu Laplce's regular terms | |
| Lee et al. | Interpretable tropical cyclone intensity estimation using Dvorak-inspired machine learning techniques | |
| CN103886335A (en) | Polarized SAR image classifying method based on fuzzy particle swarms and scattering entropy | |
| CN106022359A (en) | Fuzzy entropy space clustering analysis method based on orderly information entropy | |
| CN112241676A (en) | Method for automatically identifying terrain sundries | |
| CN103473813A (en) | Automatic extracting method for three-dimensional model members | |
| CN111611293A (en) | Outlier data mining method based on feature weighting and MapReduce | |
| CN106845538A (en) | A kind of sparse Subspace clustering method for declining optimization based on selective coordinate | |
| CN103942779A (en) | Image segmentation method based on combination of graph theory and semi-supervised learning | |
| CN102855624B (en) | A kind of image partition method based on broad sense data fields and Ncut algorithm | |
| CN106485686A (en) | One kind is based on gravitational spectral clustering image segmentation algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170613 |