CN103559716A - Method for automatic segmentation of defective image - Google Patents

Abstract

The invention discloses a method for automatic segmentation of a defective image. The method comprises the steps that A, image data are clustered through two-layer subtraction clustering; B, arrangement is conducted on center sets from big to small; C, a central point set V is initialized through first c elements; D, an efficiency analysis index of clustering is calculated; E, c=c+1 and an iterative computation is conducted until c>cmax; F, the value of V is determined when FXB (U,V,c) is the least value; G, U is calculated again through the value of V and image segmentation is conducted according to the formula that uik=max{u1k, u2k,...,uck} and xi belongs to the ith type. The method can overcome the shortages of the prior art, automatic segmentation in random walking mode is achieved, the image processing quality is improved, operation time is shortened, and the work efficiency is improved.

Description

A kind of automatic division method of defect image

Technical field

The present invention relates to graph and image processing, pattern-recognition and machine vision technique field, especially a kind of automatic division method of defect image.

Background technology

Along with widespread use and the fast development of computer technology, graph and image processing, pattern-recognition and machine vision progressively develop into popular research topic.Up to the present, researchist has proposed thousands of kinds of image partition methods, these methods are all obtaining certain achievement in varying degrees, but most achievement in research all designs for a certain types of image or for a certain concrete application background, there is no unified solution, want make concrete analyses of concrete problems, conventionally need to combine and could more effectively solve such image segmentation problem with association area knowledge, owing to lacking a unified technological means, know how according to image, to carry out design and the selection of appropriate method.Therefore, image cut apart remain image process with machine vision in the challenging difficult problem of tool, restricted greatly development and the application of this subject.

Image is cut apart technology and the process that exactly image is divided into the region of each tool characteristic and proposes interesting target.The characteristic here can be gray scale, color, texture of pixel etc., and predefined target can be to single region, also can corresponding a plurality of regions.It is to process by image the committed step that proceeds to graphical analysis that image is cut apart, and occupies critical role in Image Engineering.

Summary of the invention

The technical problem to be solved in the present invention is to provide a kind of automatic division method of defect image, can solve the deficiencies in the prior art, the segmentation effect that the Seed Points that the method provides carries out is more satisfactory, realized the auto Segmentation of random walk, improve quality of image processing, reduce the running time, improved work efficiency.

For solving the problems of the technologies described above, the technical solution used in the present invention is as follows.

An automatic division method for defect image, comprises the following steps:

A, utilize two layers of subtractive clustering to carry out cluster to view data, obtain a N _psubset;

B ,Dui center collection

arrange from big to small, and make initial value c=2;

C, with

in before c element Initialization Center point set V, utilize center

to v _idistance

degree of membership matrix u _ik, a c cluster centre v _iand objective function J _m(U, V) carries out cluster, wherein,

$d_{\mathrm{il}}^2(x_l,v_i)\≈d_{\mathrm{il}}^2(x_k^c,v_i)={\left|\right|x_k^c-v_i\left|\right|}_A^2,u_{\mathrm{ik}}={\left[{\mathrm{\Σ}}_{j=1}^c{\left[\frac{d_{\mathrm{ik}}(x_k,v_i)}{d_{\mathrm{jk}}(x_k,v_j)}\right]}^{2/(m-1)}\right]}^{-1},$

$v_i=\frac{{\mathrm{\Σ}}_{k=1}^{N_p}{\left(u_{\mathrm{ik}}\right)}^m{x}_k^c{D}^c\left(x_k^c\right)}{{\mathrm{\Σ}}_{k=1}^{N_p}{\left(u_{\mathrm{ik}}\right)}^m{D}^c\left(x_k^c\right)},J_m(U,V)={\mathrm{\Σ}}_{k=1}^{N_p}{\mathrm{\Σ}}_{i=1}^c{\left(u_{\mathrm{ik}}\right)}^m{d}_{\mathrm{ik}}^2(x_k^c,v_i)D^c\left(x_k^c\right),$

k＝1，2，...，n；

D, calculating Clustering Validity Analysis index:

$F_{\mathrm{XB}}(U,V,c)=\frac{J_m(U,V)}{n\left(\min \right||v_i-v_j|\left|\right)}i,j=\mathrm{1,2},...,c;$

E, make c=c+1, if c > is c _max, turn to step F; Otherwise, turn to step C;

F, definite $\underset{V}{\inf }\{V=\underset{i,j=\mathrm{1,2},...,c}{\arg \min }\frac{J_m(U,V)}{n\left(\min \right||v_i-v_j|\left|\right)}\},$ Be F _xBv value when (U, V, c) gets minimum value;

G, utilize V to recalculate U and according to u _ik=max{u _1k, u _2k..., u _ckx _i∈ i class is carried out image and is cut apart.

As preferably, in steps A, image is removed to noise processed.

As preferably, in steps A, using the pixel of image as sample point.

The present invention combines fast fuzzy C-means clustering with traditional random walk method, utilize quick FCM to cut apart Seed Points number and position that obtained cluster numbers C and corresponding cluster centre are determined random walk, the random walk method of the Automatic Logos based on fast fuzzy C-means clustering has been proposed, the method of proposition is applied to cutting apart of road defect image, experimental result shows, the segmentation effect that the Seed Points that the method provides carries out is more satisfactory, realized the auto Segmentation of random walk, reduce the running time, improved work efficiency.

First utilize quick FCM to Image Segmentation Using, obtain cluster numbers and cluster centre point coordinate; Then take this information chooses priori (being selected seed point) is provided as the Seed Points of random walk method; Finally, quick FCM is fused in random walk dividing method, makes random walk can realize auto Segmentation.

Fast fuzzy C-means clustering carries out image to be cut apart, and is to using the pixel of image as sample point, degree of membership u _ikthat the possibility that belongs to a certain kinds to pixel is directly proportional, and this possibility only depends on the distance at pixel and each class center, when near the pixel corresponding class center is assigned to compared with high degree of membership value and while being assigned to lower degree of membership value away from the pixel at class center, in class, error sum of squares reaches global minimum.

Random walk method, first carries out pre-service to image, removes noise; Then image is regarded as to figure, and calculate the weights between each point according to weight function, according to quick FCM, cut apart and obtain the Seed Points that cluster numbers and cluster centre coordinate are determined target and background, random walk on figure, until all points are all absorbed, according to each point, arrive the probability of Seed Points to the some classification on figure, and Output rusults.

Adopt the beneficial effect that technique scheme is brought to be: the present invention to be cut apart and the image pre-service such as medical image for road defect image, there is good recognition efficiency and accuracy, and can improve image processing speed.

Accompanying drawing explanation

Fig. 1 is the entire flow figure that defect image is processed.

Fig. 2 is the process flow diagram of quick FCM dividing processing.

Fig. 3 is the process flow diagram that random walk is processed.

Fig. 4 is the original image of ship image in embodiment.

Fig. 5 is the image of ship image after existing dividing method is processed in embodiment.

Fig. 6 is the image of ship image after dividing method provided by the invention is processed in embodiment.

Fig. 7 is the original image of road transverse crack image in embodiment.

Fig. 8 is the image of road transverse crack image after existing dividing method is processed in embodiment.

Fig. 9 is the image of road transverse crack image after dividing method provided by the invention is processed in embodiment.

Embodiment

An automatic division method for defect image, comprises the following steps:

A, utilize two layers of subtractive clustering to carry out cluster to view data, obtain a N _psubset;

B ,Dui center collection

arrange from big to small, and make initial value c=2;

C, with

in before c element Initialization Center point set V, utilize center

to v _idistance

degree of membership matrix u _ik, a c cluster centre v _iand objective function J _m(U, V) carries out cluster, wherein,

$d_{\mathrm{il}}^2(x_l,v_i)\≈d_{\mathrm{il}}^2(x_k^c,v_i)={\left|\right|x_k^c-v_i\left|\right|}_A^2,u_{\mathrm{ik}}={\left[{\mathrm{\Σ}}_{j=1}^c{\left[\frac{d_{\mathrm{ik}}(x_k,v_i)}{d_{\mathrm{jk}}(x_k,v_j)}\right]}^{2/(m-1)}\right]}^{-1},$

$v_i=\frac{{\mathrm{\Σ}}_{k=1}^{N_p}{\left(u_{\mathrm{ik}}\right)}^m{x}_k^c{D}^c\left(x_k^c\right)}{{\mathrm{\Σ}}_{k=1}^{N_p}{\left(u_{\mathrm{ik}}\right)}^m{D}^c\left(x_k^c\right)},J_m(U,V)={\mathrm{\Σ}}_{k=1}^{N_p}{\mathrm{\Σ}}_{i=1}^c{\left(u_{\mathrm{ik}}\right)}^m{d}_{\mathrm{ik}}^2(x_k^c,v_i)D^c\left(x_k^c\right),$

k＝1，2，...，n；

D, calculating Clustering Validity Analysis index:

$F_{\mathrm{XB}}(U,V,c)=\frac{J_m(U,V)}{n\left(\min \right||v_i-v_j|\left|\right)}i,j=\mathrm{1,2},...,c;$

E, make c=c+1, if c > is c _max, turn to step F; Otherwise, turn to step C;

F, definite $\underset{V}{\inf }\{V=\underset{i,j=\mathrm{1,2},...,c}{\arg \min }\frac{J_m(U,V)}{n\left(\min \right||v_i-v_j|\left|\right)}\},$ Be F _xBv value when (U, V, c) gets minimum value;

G, utilize V to recalculate U and according to u _ik=max{u _1k, u _2k..., u _ckx _i∈ i class is carried out image and is cut apart.

It should be noted that in steps A, with two layers of subtractive clustering, determine N _pthe sample set that individual image color is close.In ground floor, by data set C _xbe equally divided into the subset that t number is b (n=bt), be designated as

data point x _idensity function be defined as the data point number in this vertex neighborhood, x _ineighborhood be defined as with x _icentered by, the suprasphere that r is radius, r must get a suitable numerical value.Defined function

$u\left(x\right)=\left\{\begin{array}{cc}1& x\&GreaterEqual;0\\ 0& x<0\end{array}\right.$

?

data point density in neighborhood is

$D_i^l\left(x_i^l\right)={\mathrm{\&Sigma;}}_{j=1}^{b}u(r-||x_i^l-x_j^l|\left|\right)$

If data point

there is maximum distribution density,

subset so

the 1st cluster centre be

$c_{c1}^l=\frac{{\mathrm{\&Sigma;}}_{x_j^l\&Element;C_{c1}^l}{x}_j^l{D}_j^l\left(x_j^l\right)}{{\mathrm{\&Sigma;}}_{x_j^l\&Element;C_{c1}^l}{D}_j^l\left(x_j^l\right)}$

Wherein, for

data point set in neighborhood.Find out after the 1st class cluster centre, then in set search the data point with maximal density functional value

and the barycenter of the data point in its neighborhood is as the 2nd cluster centre

repeat said process, until

the Center Number of k for obtaining.Profit uses the same method the data set upgrading is carried out to second layer subtractive clustering, and density function is $D_i={\mathrm{\&Sigma;}}_{j=1}^{N_c}u(r-||x_i^l-x_j^l|\left|\right)D^1\left(x_j\right),$ N wherein _cand D ¹(x _j) be Center Number and the center density of ground floor, therefore, can obtain cluster centre c _cj, j=1,2 ..., N _p, N _pcenter Number for second layer acquisition.。

It should be noted that in step C, utilize hierarchical subtractive clustering method, according to similarity criterion, image point set S is divided into N _pindividual subset S _k(k=1,2 ..., N _p), consider each subset S _kthe color of interior pixel is more approaching,

to central point v _idistance can use approx S _kcenter to v _idistance

represent, $d_{\mathrm{il}}^2(x_l,v_i)\&ap;d_{\mathrm{il}}^2(x_k^c,v_i)={\left|\right|x_k^c-v_i\left|\right|}_A^2;$

Membership function u _ik(represent k sample x _kwith i cluster centre v _imembership, i.e. x _kbelong to v _idegree) be: $u_{\mathrm{ik}}={\left[{\mathrm{\&Sigma;}}_{j=1}^c{\left[\frac{d_{\mathrm{ik}}(x_k,v_i)}{d_{\mathrm{jk}}(x_k,v_j)}\right]}^{2/(m-1)}\right]}^{-1}$

The size of fuzzy matrix U becomes N from original n * c _p* c, c cluster centre v _iand objective function J _m(U, V) is respectively

$v_i=\frac{{\mathrm{\&Sigma;}}_{k=1}^{N_p}{\left(u_{\mathrm{ik}}\right)}^m{x}_k^c{D}^c\left(x_k^c\right)}{{\mathrm{\&Sigma;}}_{k=1}^{N_p}{\left(u_{\mathrm{ik}}\right)}^m{D}^c\left(x_k^c\right)},J_m(U,V)={\mathrm{\&Sigma;}}_{k=1}^{N_p}{\mathrm{\&Sigma;}}_{i=1}^c{\left(u_{\mathrm{ik}}\right)}^m{d}_{\mathrm{ik}}^2(x_k^c,v_i)D^c\left(x_k^c\right),$

k＝1，2，...，n。

The Matrix dividing of fuzzy clustering all belongs to following division space:

$M_{\mathrm{fc}}=\{U\&Element;R^{\mathrm{cn}}|u_{\mathrm{ik}}\&Element;\left[\mathrm{0,1}\right],\&ForAll;i,k;{\mathrm{\&Sigma;}}_{i=1}^c{u}_{\mathrm{ik}}=1,\&ForAll;k;0<{\mathrm{\&Sigma;}}_{k=1}^n{u}_{\mathrm{ik}}<n,\&ForAll;i\}$

M ∈ [1, ∞) be FUZZY WEIGHTED index, be called again smoothing factor, on the one hand, it affects the concavity and convexity of objective function, controls again on the other hand the ambiguity of cluster, is controlling sample and share degree between fuzzy class.

As preferably, in steps A, image is removed to noise processed.

As preferably, in steps A, using the pixel of image as sample point.

By method of the present invention, random walk can be realized defect image auto Segmentation.

As shown in Fig. 6 and Fig. 9, by using the method providing in this embodiment to process the image of ship and crack on road, can find out in image that ship and crack are successfully split.

Take MatlabR2013a as experiment porch, and the Data Comparison in processing time is as follows:

More than show and described ultimate principle of the present invention and principal character and advantage of the present invention.The technician of the industry should understand; the present invention is not restricted to the described embodiments; that in above-described embodiment and instructions, describes just illustrates principle of the present invention; without departing from the spirit and scope of the present invention; the present invention also has various changes and modifications, and these changes and improvements all fall in the claimed scope of the invention.The claimed scope of the present invention is defined by appending claims and equivalent thereof.

Claims (3)

1. an automatic division method for defect image, is characterized in that comprising the following steps:

A, utilize two layers of subtractive clustering to carry out cluster to view data, obtain a N _psubset;

B ,Dui center collection

arrange from big to small, and make initial value c=2;

C, with

in before c element Initialization Center point set V, utilize center

to v _idistance

degree of membership matrix u _ik, a c cluster centre v _iand objective function J _m(U, V) carries out cluster, wherein,

$d_{\mathrm{il}}^2(x_l,v_i)\&ap;d_{\mathrm{il}}(x_k^c,v_i)={\left|\right|x_k^c-v_i\left|\right|}_A^2,u_{\mathrm{ik}}={\left[{\mathrm{\&Sigma;}}_{j=1}^c{\left[\frac{d_{\mathrm{ik}}(x_k,v_i)}{d_{\mathrm{jk}}(x_k,v_j)}\right]}^{2/(m-1)}\right]}^{-1},$

$v_i=\frac{{\mathrm{\&Sigma;}}_{k=1}^{N_p}{\left(u_{\mathrm{ik}}\right)}^m{x}_k^c{D}^c\left(x_k^c\right)}{{\mathrm{\&Sigma;}}_{k=1}^{N_p}{\left(u_{\mathrm{ik}}\right)}^m{D}^c\left(x_k^c\right)},J_m(U,V)={\mathrm{\&Sigma;}}_{k=1}^{N_p}{\mathrm{\&Sigma;}}_{i=1}^c{\left(u_{\mathrm{ik}}\right)}^m{d}_{\mathrm{ik}}^2(x_k^c,v_i)D^c\left(x_k^c\right),$

k＝1，2，...，n；

D, calculating Clustering Validity Analysis index:

$F_{\mathrm{XB}}(U,V,c)=\frac{J_m(U,V)}{n\left(\min \right||v_i-v_j|\left|\right)}i,j=\mathrm{1,2},...,c;$

E, make c=c+1, if c > is c _max, turn to step F; Otherwise, turn to step C;

F, definite $\underset{V}{\inf }\{V=\underset{i,j=\mathrm{1,2},...,c}{\arg \min }\frac{J_m(U,V)}{n\left(\min \right||v_i-v_j|\left|\right)}\},$ Be F _xBv value when (U, V, c) gets minimum value;

G, utilize V to recalculate U and according to u _ik=max{u _1k, u _2k..., u _ckx _i∈ i class is carried out image and is cut apart.

2. the automatic division method of defect image according to claim 1, is characterized in that: in steps A, image is removed to noise processed.

3. the automatic division method of defect image according to claim 2, is characterized in that: in steps A, using the pixel of image as sample point.

Images