CN101840577B - Image automatic segmentation method based on graph cut - Google Patents

Abstract

The invention discloses an automatic segmentation method based on graph cut for color images and gray level images, mainly solving the problems of the existing graph cut technology that interaction and modeling are required in graph cut and the segmentation result is required to be modified manually. The method comprises the following steps: dividing an image into an inner area and an outer area; establishing the data item of the energy function according to the similarity of pixels in different areas, wherein mean shift, YCbCr color space conversion and block partition are adopted in calculation of the similarity; establishing the smoothing item of the energy function according to the marginal information and spatial location of the image; adopting graph cut to perform optimization to the energy function, thus realizing one-step cutting to the image; and using the segmentation result as the new inner and outer areas, performing iterative execution of the above operations, and stopping iterative execution when iterative conditions are satisfied. The method has the advantages of automation, good effect and less iterations and can be used in the computer vision fields such as image processing, image editing, image classification, image identification and the like.

Description

Image automatic segmentation method based on the figure cutting

Technical field

The present invention relates to image processing field, a kind of method of image segmentation specifically, this method can effectively be cut apart coloured image and gray level image, is divided into two parts of target and background, can be subsequent image processing, identification, classification etc. the basis is provided.

Background technology

Image segmentation is a basic and crucial problem in image processing field and the computer vision.Its purpose of cutting apart extracts the interested target of people exactly from image background, for follow-up analysis, understanding, classification, tracking, identification and processing etc. provide the basis.The application of image segmentation is very extensive, and for example: medical science, Flame Image Process, military affairs, physical culture, intelligent transportation, industrial or agricultural etc. almost are applied in all spectra of relevant Flame Image Process.Because the importance of image segmentation just has great importance to Study of Image Segmentation.

The algorithm of image segmentation is a lot, but does not up to the present also have a method in common, remains a challenging research direction.Classification to it also is varied, can obtain different classification results from different angles.

Information according to image is classified, and can the method for image segmentation be divided into: based on the dividing method of area information, based on the dividing method of marginal information and the method for calmodulin binding domain CaM information and marginal information.Lay particular emphasis on based on the method in zone and to utilize the similarity of image-region internal feature that image is cut apart.Mainly based on the uncontinuity of image gray levels, realize cutting apart based on the cutting techniques at edge to image through the border of detecting between the different homogeneous areas.

Degree according to user in image segmentation participates in can be divided into interactively image segmentation and automatic image segmentation with image partition method.

Various technology according to image partition method itself is used are classified, and can be divided into: based on the method for cluster, based on morphology methods, based on neural network method, based on the method for fuzzy set, based on method of wavelet, based on the method for genetic algorithm, based on fractal geometry and based on graph theory method etc.

Wherein the image segmentation based on Graph-theoretical Approach is a kind of new technology, new construction; But the relation of good treatment overall situation and partial situation; And divisiblely going out good result, cause the interest that People more and more is many in recent years, is the new research focus in image segmentation field in the world; Become accurate and useful instrument day by day, and obtained successful application.Image segmentation based on graph theory is exactly the network chart that represents the image as cum rights, and the pixel of the node correspondence image of network chart through on network chart, carrying out sequence of operations and computing, is accomplished cutting apart image.

Numerous based on graph theory method in, the figure cutting technique is energy-optimised and especially noticeable with it, in recent years, it is employed some problems that solve in computer vision, computer graphics and the machine learning.Wherein Boykov et.al. etc. is a series of problems in the computer vision, as image recovery, stereoscopic vision, many viewpoints rebuild, target is cut apart etc., and problem is converted into finds the solution energy-minimum, thereby use figure cutting technique solves these problems effectively.In fact, above-mentioned these problems all can classify as the label problem, so just are converted into energy function optimization looking for most probable label, and the method that converts the energy function optimization problem into figure cutting again solves.

Method based on the figure cutting has extraordinary framework, can fully utilize various characteristic informations, adopts the various thoughts of cutting apart; Incorporate in the multiple cutting techniques; And can exempt because the error that discretize causes obtains accurately result, so in image segmentation, can be used widely.

Calendar year 2001; Boykov Y; People such as Jolly M P have delivered one piece of classical documents (Interactive graph cutsfor optimal boundary and region segmentation of objects in n-d images.In:Proc.IEEEInternational Conference on Computer Vision.2001; Pp.105-112), the document has caused the upsurge of figure cutting technique in image segmentation.This piece n dimension image Interactive Segmentation in this piece article, formed the basis of image segmentation alternately.The user provides constraint through being sampled as of prospect and background seed points cut apart; With histogram model view data is estimated; Energy function is realized minimizing having set up on the energy function model based use figure cutting algorithm, thereby reach the purpose of image segmentation.This method is with its simple and direct interactive mode, processing speed faster, and can various information fusion be caused people's extensive concern.

Rother C subsequently; Kolmogorov V; People such as Blake A propose the Grabcut method on the basis of Boykov et.al., document is GrabCut:interactive foreground extraction using iterated graph cuts.ACM Transactions on Graphics (TOG), and 2004; 23 (3), pp.309～314.Grab Cut has done the improvement aspect three on the basis of figure cutting: the histogram model of gray level image is replaced by the gauss hybrid models of coloured image, and coloured image is cut apart; More powerful iterative process substituting disposable minimal cut can be between estimation and parameter learning alternately; Mutual selection becomes simple, loosens, and allows not exclusively to demarcate.The user only needs two of rectangle frame of specific context to angle point.The Grab cut of iteration, more succinct on user's interactive mode, be a kind of image partition method best in these class methods at present.

The deficiency that the interactive image dividing method that cuts based on figure at present exists is:

1) needs the user to participate in,, under certain constraints, accomplish the cutting apart of image based on the estimation of data model, and can not realize cutting apart automatically through algorithm to image through the appointment seed points of user interactions;

2) tie up in the mutual image partition method at the n of Boykov, the histogram probability model can approach any probability distribution under certain precision, and still along with the raising of precision, must guarantee has a large amount of number of individuals samplings.And for interactively Grab cut method, the selection of the hybrid parameter of gauss hybrid models has very big influence to cutting apart, and how to select this parameter, does not also have better scheme at present;

3) n of Boykov ties up mutual image partition method and cuts apart to gray level image, and Grab cut method is cut apart to coloured image;

The distribution estimating model of the choosing of quality of 4) cutting apart and seed points, data all has very big relation, and inappropriate seed points can't obtain correct result, and choosing according to different people, different situations of seed points all has very big otherness.And Interactive Segmentation often can not obtain good segmentation result fast, also needs the user that the result of cutting apart is carried out further revising meticulously, just can obtain gratifying result.

Summary of the invention

The objective of the invention is to overcome the deficiency of prior art; A kind of automated graphics dividing method based on the figure cutting is proposed; The distribution of data is estimated need not to set up under the condition of data model; Realization is carried out cutting apart automatically of target and background to coloured image and gray level image, and obtains good segmentation result with fast speeds.

The technical thought that realizes the object of the invention is through image division being become two zoness of different, and the smooth item in the data item of setting up energy function and the energy function is also found the solution, and realizes cutting apart of image automatically.Concrete performing step comprises as follows:

(1) the input field of definition is the image I to be split of Ω, and the initialization of cutting apart makes it form interior zone C _iWith perimeter C _o, Ω=C _i∪ C _o

(2) treat split image I and carry out filtering, filtered image is designated as I _mFiltered image is carried out color space conversion, and the image after the color conversion is designated as I _cImage after the color conversion is carried out piece divide, the image after piece is divided is designated as I _b

(3) image I after the calculation of filtered _mIn each pixel and piece partitioned image I _bInterior zone C _iIn the similarity sum of each piece pixel:

And and perimeter C _oIn the similarity sum of each piece pixel:

And with it as data item,

Wherein, P is I _mThe set of pixels of image, m, e are pixels among the P, C _iBe I _bInterior zone, C _oBe I _bThe perimeter, n is C _iIn the piece pixel, f is C _oIn the piece pixel, S (m, the n) similarity between pixel n in remarked pixel m and the interior zone, S (e, f) similarity between pixel f in remarked pixel e and the perimeter;

(4) set up smooth according to image edge information and spatial relation:

V _p，q(L _p，L _q)＝V(L _p，L _q)*w _pq，

Wherein, w _PqBe the VG (vertical gradient) information and the horizontal gradient information at pixel edge, the variation of reflection pixel space, V (L _p, L _q) relation of reflection pixel locus, definition as follows:

$V(L_p,L_q)=\left\{\begin{array}{cc}1,& L_p\≠L_q\\ 0,& L_p=L_q\end{array}\right.;$

(5) according to smooth in the data item in (3) and (4), set up energy function E:

$E\left(L\right)=\underset{m\∈P}{\mathrm{\Σ}}{D}_m\left({{C_{\′}}_i}^{\′}\right)+\underset{e\∈P}{\mathrm{\Σ}}{D}_e\left({{C_{\′}}_o}^{\′}\right)+\underset{p,q\∈N}{\mathrm{\Σ}}{V}_{p,q}(L_p,L_q);$

(6) energy function of setting up is found the solution, obtain the result of cutting apart first of image I;

(7) will be first target in the segmentation result as interior zone, background is as the perimeter, repeated execution of steps (3) is exported final segmentation result until satisfying the iterated conditional of setting.

The present invention has following advantage:

(1) the present invention is automatic completion owing to the initialization of image is cut apart; And the calculating, smooth calculating, finding the solution all of energy function of data item realizes based on algorithm automatically in the energy function; Need not any participation of user, thereby can realize cutting apart automatically image;

(2) the present invention is owing to adopt the similitude between pixel to the calculating of data item in the energy function; Smooth calculating is based on image edge information and spatial relation; Need not any priori and the distribution of data is estimated, exempted the foundation of view data model and the estimation and the study of relevant parameter;

(3) the present invention is because in the cutting apart of image, consideration be the similarity between image pixel, need not to consider is to the gray level image modeling or to the coloured image modeling, thus this invention not only can cut apart gray level image, also can be to color images;

(4) the present invention finds the solution owing to energy function employing figure is cut, and can realize the optimum segmentation of binary; Adopt filtering owing to treat split image simultaneously, made it become level and smooth; Because filtered image is carried out color space conversion, have the better cluster effect in addition, these two kinds of processing make that all the calculating of similarity is more accurate;

(5) the present invention is because the dividing method of figure cutting is based on the dividing method in zone; And these class methods tend to cause over-segmentation; Image segmentation is become too much zone; If promptly in based on the structure in zone, do not consider the marginal information of image object, may cause noise margin or object inside cavitation to occur, so in smooth of energy function, considered image edge information; Utilize the gradient information of image, thereby can improve the quality of image segmentation in level and vertical direction;

(6) the present invention is in the process of cutting apart; Divide processing because the image after the color conversion has been carried out piece, make the calculating of similarity become fast, and owing to need not model parameter is estimated and study; Can use considerably less iterations, make splitting speed be improved.

Description of drawings

Fig. 1 is a FB(flow block) of the present invention;

Fig. 2 is the initialization synoptic diagram of the present invention to image;

The synoptic diagram that Fig. 3 divides image block for the present invention;

Fig. 4 is the synoptic diagram of figure cutting of the present invention;

Fig. 5 is cut apart analogous diagram for the present invention to coloured image automatically;

Fig. 6 is cut apart analogous diagram for the present invention to gray level image automatically.

Embodiment

Followingly the present invention is described in further detail with reference to accompanying drawing.

With reference to accompanying drawing 1, performing step of the present invention comprises as follows:

Step 1: treat the cutting object initialization.

(1a) import image I to be split; And the line number of establishing this image is r, and columns is c, being the center of circle by the center of split image; With min (r/6; C/6) be radius, on image to be split, constitute a closed contour curve C, the closed contour curve C is that the image division of Ω is inside and outside two zone: C with field of definition _iAnd C _o, Ω=C _i∪ C _o∪ C;

(1b) pixel in first closure curve C and the C interior zone is classified as one type, is designated as C _iPixel in the perimeter of curve C is classified as one type, is designated as C _o, as shown in Figure 2, wherein white portion belongs to one type, expression interior zone C _i, black region belongs to one type, expression perimeter C _o

Step 2: treat conversion and piece division that split image I carries out filtering, color space.

(2a) treat split image I and adopt average drifting Mean Shift to carry out Filtering Processing, average drifting not only can carry out filtering to image, and does not lose edge of image property, and the image after average drifting is handled is designated as I _m

(2b) pass through following column matrix to the RGB image I _mCarry out the conversion of YCbCr color space, be designated as Ic, and obtain Y, Cb and three independent components of Cr.

$\left[\begin{array}{c}Y\\ \mathrm{Cb}\\ \mathrm{Cr}\end{array}\right]=\left[\begin{array}{ccc}0.299& 0.587& 0.114\\ -0.1687& -0.3313& 0.500\\ 0.500& -0.4187& -0.0813\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\end{array}\right]+\left[\begin{array}{c}0\\ 128\\ 128\end{array}\right]$

In the matrix, Y is brightness; Cb and Cr are respectively the aberration between blueness and danger signal and the luminance signal, in color component, have rejected the brightness composition, and the luminance component that is about in the color separates;

(2c) to I _cImage carries out the division of piece, is divided into n * m piece, and the size of each piece is 15 * 15, m, and the calculating of n is following:

$n=\left(\frac{r-1}{15}\right)+1,$

$m=\left(\frac{c-1}{15}\right)+1.$

Wherein r is the line number of image to be split, and c is the columns of image to be split,

Image after piece is divided is designated as I _b, as shown in Figure 3, I _bIn the characteristic of each piece pixel from the characteristic average of corresponding 15 * 15 pixels in the Ic image, i.e. I _bIn the color component S of any element j _jBe image I _cThe average of all 15 * 15 color components in pixels in the piece, S _jComputing formula as shown in the formula:

$S_j=\frac{1}{15*15}\underset{s}{\mathrm{\Σ}}{I}_c\left(T_i^s\right)$

is certain all pixel of 15 * 15 among the Ic.

Step 3: computational data item.

(3a) with two zone C _iWith C _oBetween the dissimilarity problem convert C into _iWith C _oSimilarity problem in the same area:

Like accompanying drawing 2, curved profile C is divided into inside and outside two parts with image, and interior zone is expressed as C _i, the perimeter is expressed as C _o, pixel v _o∈ C _o, pixel v _i∈ C _i, w (v _o, v _i) remarked pixel v _o, v _iBetween power, this power has reflected the similarity between two pixels, and the weights between two pixels are more little, then two pixels are similar more.The target that use figure cutting is cut apart is to use the minimal cut criterion, and image is cut into two zones, make the weights and the minimum on the limit that connects these two zones promptly to have minimum similarity between these two zones, the minimal cut criterion as shown in the formula:

$\underset{C}{\mathrm{Min}}C=\underset{v_o\∈C_o,v_i\∈C_i}{\mathrm{\Σ}}w(v_o,v_i)$

The present invention converts two interregional dissimilarity problems between the same area similarity problem when utilizing the minimal cut criterion of figure cutting, the weights sum is more little in the following formula, represent two interregional more dissimilar, be converted into C _iWith C _oThe similarity problem of interior zone, i.e. C separately _iWith C _oCertain characteristic between each pixel of interior zone and the pixel is more little like the gray-scale value difference separately, and is similar more, and the formula after the conversion is:

$\underset{C}{\mathrm{Min}}C=\underset{e,f\∈C_o}{\mathrm{\Σ}}S(e,f)+\underset{m,n\∈C_i}{\mathrm{\Σ}}S(m,n)$

(a, b) distinctiveness of characteristic such as gray-scale value between remarked pixel a and the b is called similarity to S, the objective of the invention is to ask for a closed curve C, and image is divided into inside and outside two zone: C _iWith C _o, having strong similarity between each regional interior pixel in these two zones, promptly very little difference thinks that then these pixels belong to same type;

(3b) computing formula of definition similarity:

If S (i, the j) similarity of j pixel among i pixel and the image T ' among the presentation video T, then computing formula as shown in the formula:

$S(i,j)=\sqrt{{(T_{\mathrm{iy}}-{T^{\′}}_{\mathrm{jy}})}^2+{(T_{\mathrm{icb}}-{T^{\′}}_{\mathrm{jcb}})}^2+{(T_{\mathrm{icr}}-{T^{\′}}_{\mathrm{jcr}})}^2}$

T wherein _IyThe luminance component of remarked pixel i, T _IcbAnd T _IcrThe color difference components of remarked pixel i; T ' _JyThe luminance component of remarked pixel j, T ' _JcbAnd T ' _JcrThe color difference components of remarked pixel j.

(3c) with the initialization segmentation result with I _bImage division becomes inside and outside two zones: interior zone C _iWith perimeter C _o, the image I after the first calculation of filtered _mIn each pixel and piece partitioned image I _bInterior zone C _iIn the similarity sum of each piece pixel:

$D_m\left({{C_{\′}}_i}^{\′}\right)=\underset{m\∈P,n\∈C_i}{\mathrm{\Σ}}S(m,n),$

Image I after the calculation of filtered again _mIn each pixel and piece partitioned image I _bPerimeter C _oIn the similarity sum of each piece pixel:

$D_e\left({{C_{\′}}_o}^{\′}\right)=\underset{e\∈P,f\∈C_o}{\mathrm{\Σ}}S(e,f),$

And with this D _m(' C _i) and De (' C _o') as data item,

Wherein, P is I _mThe set of pixels of image, m, e are pixels among the P, C _iBe I _bInterior zone, C _oBe I _bThe perimeter, n is C _iIn the piece pixel, f is C _oIn the piece pixel, S (m, the n) similarity between pixel n in remarked pixel m and the interior zone, S (e, f) similarity between pixel f in remarked pixel e and the perimeter.

Step 4: calculate smooth.

(4a) establish any two neighbors that p and q are images, Lp is the pixel p label corresponding with q with Lq, definition V (L _p, L _q) be:

$V(L_p,L_q)=\left\{\begin{array}{cc}1,& L_p\≠L_q\\ 0,& L_p=L_q\end{array}\right.,$

This V (L _p, L _q) reflection image pixel spatial relationship;

(4b) use Gaussian filter to treat split image earlier and carry out smoothing processing, utilize the Sobel operator that the image after level and smooth is carried out rim detection again, obtain image border to be split horizontal gradient information hC (r, c) and VG (vertical gradient) information vC (r; C), according to this horizontal gradient information hC (r, c) and VG (vertical gradient) information vC (r, c); Judge the variation of image pixel, if p=(r, c); Q=(r+1, c), w then _Pq=vC (r, c); If p=(r, c), q=(r, c+1), w then _Pq=hC (r, c),

Wherein, p=(r, the c) pixel of the capable c row of expression r, q=(r+1, the c) pixel of the capable c row of expression r+1, q=(r, the c+1) pixel of the capable c+1 row of expression r, w _PqIt is the parameter that the reflection image pixel changes;

(4c) definition V _{P, q}(L _p, L _q) be smooth, this computing formula of smooth is following:

V _p，q(L _p，L _q)＝V(L _p，L _q)*w _pq。

Step 5: set up energy function according to data item and smooth item:

The D that adopts step 3 to calculate _m(' C _i') and De (' C _o') as data item, the V that adopts step 4 to calculate _{P, q}(L _p, L _q) as smooth, energy function then of the present invention as shown in the formula:

$E\left(L\right)=\underset{m\∈P}{\mathrm{\Σ}}{D}_m\left({{C_{\′}}_i}^{\′}\right)+\underset{e\∈P}{\mathrm{\Σ}}{D}_e\left({{C_{\′}}_o}^{\′}\right)+\underset{p,q\∈N}{\mathrm{\Σ}}{V}_{p,q}(L_p,L_q).$

Step 6: energy function is found the solution, and obtain segmentation result first.

(6a) set up network chart;

(6a1) set up the node of network chart, { wherein v is corresponding to each pixel of image for s, t}, and s and t are two extra nodes, and s is called the source, and t is called remittance for node set V={v} ∪;

(6a2) set up the limit of network chart, limit set E={e} ∪ v, s}, v, t}, wherein e} is that the n-of network chart connects, be each node v with its neighborhood node between be connected; V, s}, v, t}} are that the t-of network chart connects, and be each node with terminal point s and t between be connected;

(6a3) the data item D of use energy function _m(' C _i') and De (' C _o') power that connects of computational grid figure t-, use smooth V of energy function _{P, q}(L _p, L _q) power that connects of computational grid figure n-, the weights such as the table 1 of network chart fillet:

The weights of network chart among table 1 the present invention

The limit	Power	Remarks
			{p，q}	V p，q(L p，L q)	{p，q}∈N
{m，s}	λ·D m(′C i′)	m∈P
			{e，t}	λ·De(′C o′)	e∈P

In the table 1, λ gets the value relevant with the inside and outside region area, as shown in the formula calculating:

$\mathrm{\λ}=\frac{{\mathrm{Ac}}_n}{S_I}$

Wherein, Ac _nThe area of the interior zone of expression curve C or the area of perimeter; S _IThe size of presentation video equals the line number * columns of image I to be split;

(V E), is depicted as the network chart of foundation like accompanying drawing 4 (a) (6a4) to set up network chart G=by the node of the foundation of (6a1), limit that (6a2) sets up and the power (6a3) calculated;

(6b) network chart that (6a) sets up is asked minimal cut according to max-flow/minimal cut criterion;

(6c) network chart is asked for after the minimal cut, all pixel segmentation of image are become two parts: target and background, obtain the result that image is cut apart first, be depicted as the synoptic diagram of figure cutting like accompanying drawing 4 (b).

Step 7: iteration is carried out.

(7a) will be first target in the segmentation result as interior zone, background is as the perimeter, it is regional to obtain new inside and outside;

The new interior exterior domain that (7b) produces for (7a) begins to repeat work from step 3;

(7c) judge whether finishing iteration work according to iterated conditional, iterated conditional can adopt two kinds of schemes: the one, and rule of thumb, iterations is set to 5 times, after 5 times iteration is carried out, finishing iteration; The 2nd, the binary segmentation result of comparison this and last time, if the pixel count that changes in this segmentation result is less than or equal to 5, finishing iteration process then;

Zone, inside and outside after (7d) iteration finishes is final binary image segmentation result, therefrom obtains the image final objective.

For verifying validity of the present invention and correctness, adopted two groups of emulation experiments, all emulation experiments all adopt Matlab 7.0 softwares to realize under Windows XP operating system.

Emulation one

Coloured image is carried out emulation experiment, and the image size that is adopted is 144 * 216 * 3, and the form of color space is RGB, and simulation process and source images are as shown in Figure 5.Wherein Fig. 5 (a) is an original image; Fig. 5 (b) is to the initialized result of original image; Fig. 5 (c) is to the filtered result of original image average drifting; Fig. 5 (d) is for carrying out the result after the YCbCr color space conversion to the image behind the average drifting, Fig. 5 (e) divides the piece that color space changes the back image, is followed successively by luminance component Y, blue color difference component C _bAnd red color component C _r, Fig. 5 (f) is the result after cutting apart for the first time, and Fig. 5 (g) is the result after cutting apart for the second time, and Fig. 5 (h) is the result after cutting apart for the third time.

Can find out that from the initialization result of Fig. 5 (b) initialization can be provided with according to the image size automatically, provide basic for realizing cutting apart automatically; Can find out that from Fig. 5 (c) original image is through after the average drifting, it is smooth that image becomes, and don't lose marginality; Result after the YCbCr color space conversion of Fig. 5 (d) can find out to have better cluster property at this color space, makes that the calculating of similarity is more accurate; Can find out that from Fig. 5 (e) image after the color conversion is carried out piece divide, each piece has identical characteristic, is that unit calculates with the piece, makes the computing velocity of similarity be improved; Can find out from the segmentation result first time of Fig. 5 (f), obtain a reasonable segmentation result basically, but some detail sections of image are also not separated fully; Can find out from Fig. 5 (g), obtain after the iteration for the second time than the better segmentation result first time; Like Fig. 5 (h), the result of cutting apart for the third time compares with the result of cutting apart for the second time, and both results are the same, satisfy iterated conditional two, explain that segmentation result does not change, so stop interative computation, the final image segmentation result that obtains satisfaction.

In whole process to coloured image emulation, there be not the mutual of user, iteration only needs 3 times, and image segmentation is not only accomplished faster automatically, and has obtained good result.

Emulation two

Gray level image is carried out emulation experiment, and the image size that is adopted is 200 * 200, and gray level is 256 looks, and simulation process and source images are as shown in Figure 6.Wherein Fig. 6 (a) is an original image; Fig. 6 (b) is an initialization result; Fig. 6 (c) is to the filtered result of original image average drifting, and Fig. 6 (d) divides the piece of image behind the average drifting, and 6 (e) figure is the result after cutting apart for the first time; Fig. 6 (f) is the result after cutting apart for the second time, and Fig. 6 (g) is the result after cutting apart for the third time.

Can find out that from the initialization result of Fig. 6 (b) initialization can be provided with according to the image size automatically, provide basic for realizing cutting apart automatically; Can find out that from Fig. 6 (c) original image is through after the average drifting, it is smooth that image becomes, and don't lose marginality; Can find out that from Fig. 6 (d) image behind the average drifting is carried out piece divide, each piece has identical characteristic, is that unit calculates with the piece, makes the computing velocity of similarity be improved; Can find out from the segmentation result first time of Fig. 6 (e), obtain a reasonable segmentation result basically, but some detail sections of image also do not split fully; Can find out from Fig. 6 (f), obtain after the iteration for the second time than the better segmentation result first time; Can find out that from Fig. 6 (g) result of cutting apart for the third time compares with the result of cutting apart for the second time, both results are the same, satisfy iterated conditional two, explain that segmentation result does not change, so stop interative computation, the final image segmentation result that obtains satisfaction.

In whole process to gray level image emulation, there be not the mutual of user equally, iteration only needs 3 times, and image segmentation is not only accomplished faster automatically, and has obtained good result, although source images is a noise image, has still obtained good segmentation result.Gray level image cuts apart that different is the conversion that need not color space with color images, so in this simulation process, there is not the result of color space conversion.

The result of these two emulation experiments has verified validity of the present invention and correctness.

Claims (2)

1. the image automatic segmentation method based on the figure cutting comprises the steps:

(1) the input field of definition is the image I to be split of Ω, and the initialization of cutting apart makes it form interior zone C _iWith perimeter C _o, Ω=C _i∪ C _o

(2) RGB image I to be split is carried out filtering, filtered image is designated as I _mFiltered RGB image is carried out the YCbCr color space conversion, and the image after the color conversion is designated as I _cImage after the color conversion is carried out piece divide, the image after piece is divided is designated as I _b

(3) image I after the calculation of filtered _mIn each pixel and piece partitioned image I _bInterior zone C _iIn the similarity sum of each piece pixel: And and perimeter C _oIn the similarity sum of each piece pixel: $D_e\left(C_o^{\′}{}_{\′}\right)=\underset{e\∈P,f\∈C_o}{\mathrm{\Σ}}S(e,f),$ And with it as data item,

Wherein, P is I _mThe set of pixels of image, m, e are pixels among the P, C _iBe I _bInterior zone, C _oBe I _bThe perimeter, n is C _iIn the piece pixel, f is C _oIn the piece pixel, S (m, the n) similarity between pixel n in remarked pixel m and the interior zone, S (e, f) similarity between pixel f in remarked pixel e and the perimeter;

(4) set up smooth according to image edge information and spatial relation:

V _p，q(L _p，L _q)＝V(L _p，L _q)*w _pq，

Wherein, w _PqBe the VG (vertical gradient) information and the horizontal gradient information at pixel edge, the variation of reflection pixel space, V (L _p, L _q) relation of reflection pixel locus, definition as follows:

$V(L_p,L_q)=\left\{\begin{array}{cc}1,& L_p\≠L_q\\ 0,& L_p=L_q\end{array}\right.;$

(5) according to smooth in the data item in (3) and (4), set up energy function E:

$E\left(L\right)=\underset{m\∈P}{\mathrm{\Σ}}{D}_m\left(C_i^{\′}{}_{\′}\right)+\underset{e\∈P}{\mathrm{\Σ}}{D}_e\left({C_o^{\′}}^{\′}\right)+\underset{p,q\∈N}{\mathrm{\Σ}}{V}_{p,q}(L_p,L_q),$

Wherein: L representes label; If p and q are any two neighbors of image, Lp is the pixel p label corresponding with q with Lq;

(6) energy function of setting up is found the solution, obtains the result of cutting apart first of image I:

(6a) multiply by the t-connection of coefficient lambda, wherein as network chart with the data item in the energy function

Ac _nThe area of the interior zone of expression curve C or the area of perimeter calculate λ D _m(' C _i') time,

Ac _iBe curve C interior zone area, calculate λ D _e(' C _o') time,

Ac _oIt is curve C perimeter area; S _IThe size of presentation video equals the line number * columns of image I to be split; With smooth V in the energy function _{P, q}(L _p, L _q) connect as the n-of network chart, set up the network chart in the graph theory;

(6b) network chart of setting up is asked minimal cut according to max-flow/minimal cut algorithm;

(7) will be first target in the segmentation result as interior zone, background is as the perimeter, repeated execution of steps (3)-step (6); Judge the iteration stopping condition, export final segmentation result: promptly rule of thumb, iterations is set to 5 times; After 5 times iteration was carried out, finishing iteration obtained final segmentation result; The perhaps binary segmentation result of this and last time relatively; If the pixel count that changes in this segmentation result is less than or equal to 5, finishing iteration process then, and with this result as final segmentation result.

2. the image automatic segmentation method based on the figure cutting according to claim 1, the wherein described initialization of cutting apart of step (1) comprises the steps:

(2a) establishing by the line number of split image is r, and columns is c, and to be the center of circle by the center of split image, (r/6 c/6) is radius, on by split image, constitutes a closed contour curve C with min;

(2b) pixel in first closure curve C and the C interior zone is classified as one type, is designated as C _iPixel in the perimeter of curve C is classified as one type, is designated as C _o

Images