CN101504768A - Color image fast partition method based on deformation contour model and graph cut - Google Patents

Abstract

The invention discloses a deformation contour model and graph cut-based method for quickly segmenting color images, which comprises the following steps: drawing a current contour line; creating a width self-adaptive annular region; constructing an s-t network for the width self-adaptive annular region; cutting at the lowest cost; determining a contour line of a segmented target; and outputting the segmented target in the image. In the method, a plane s-t network is adopted, so the construction and segmentation of the network is more convenient and quick, particularly in an iteration mode; the unidirectional expansion of the contour line can effectively reduce the superimposition of adjacent annular regions; and after the adoption of the width self-adaptive annular region, only non-standard contour line sections are segmented, and the segmentation efficiency is improved considerably. The method has the advantages of simple and convenient operation, quick segmentation speed, high precision, high noise immunity and the like, and can be used for segmenting images with various colors and grays.

Description

The color image fast partition method that cuts based on deformation contour model and figure

Technical field

The invention belongs to technical field of image processing, be specifically related to discuss the method that interesting target in the image is cut apart in conjunction with deformation contour model and figure hugger.

Background technology

Image segmentation is to utilize information such as gray scale, color, texture, shape that image segmentation is become several independently significant continuum or objects, the object that extracts after each intra-zone has the characteristic of homogeneity, cuts apart is exactly our interested target.Image segmentation is the basis of objective expression in the Image Engineering, feature extraction and parameter measurement, and then makes that more high-rise graphical analysis and understanding becomes possibility.Especially, in the medical image analysis field, image segmentation is precondition and the committed step that tissue measurement, anatomical structure analysis and tissue characterization etc. are used.

Under the condition that lacks enough prior imformations image being carried out dividing processing is a relatively difficult technologies problem, because the importance and the difficulty of this problem, the researchist is carrying out unremitting effort always both at home and abroad for many years, people utilize various mathematical theories and instrument, use different models, work out multiple image partition method, formed a huge system.But the whole bag of tricks can only design at the demand of various practical application area, and separately specific aim and limitation are arranged.Up to the present, also there is not a kind of general image dividing method.

Be man-machine interaction in the dividing method semi-automatic the cutting apart of image, finishes a kind of method of cutting apart jointly.This method can provide the effective control to cutting procedure, makes operating personnel can intervene, influence cutting procedure where necessary easily.The semi-automatic operational performance that both makes full use of computing machine of cutting apart has been brought into play people's judgment again, thereby makes and to cut apart more accurately, receives increasing concern in the image segmentation practice.

Deforming template is a kind of image partition method based on the edge commonly used.Deforming template can be regarded as the energy minimization elastic variable curve of being controlled by a different set of acting force in the image (curved surface).The definition of deforming template energy function comprises internal energy function, image energy and external energy function, the flatness requirement of deforming template is satisfied in the definition of internal energy, image energy attracts deforming template to draw close to corresponding characteristics of image (as edge, angle point), and external energy is user-defined energy, is used for satisfying certain requirement.The solution procedure of deforming template is one and seeks least energy in elastic curve distortion and motion process, makes it the process of being drawn close to feature locations by the initial position on the image gradually.

Being one based on the image segmentation of figure hugger opinion is the dividing method of optimisation technique with Markov random field as iconic model, max-flow algorithm.The core concept of figure hugger opinion is to construct an energy function, uses combined optimization technique to minimize this energy function then.That is, image segmentation problem is considered as one pixel is designated as the typical binary label combinatorial optimization problem of foreground/background, structure and the network flow theory by energy minimization model, network then is converted to the label problem with the max-flow method and solves.

The energy minimization model.For the label of calculating pixel, need energy function of structure about label.Constraint commonly used in the computer vision has data constraint and smooth constraint, and data constraint is that the brightness of corresponding point on the different images should be consistent; Smooth constraint is that the brightness of consecutive point on the same image should be similar, and it has embodied the continuity of intra-zone and the uncontinuity on border.

The structure of network.If (V E) is a non-directed graph to G=, and V is a vertex set, and E is the limit collection.For connecting x among the vertex set V, the limit e of y can be considered from x to y and two different directions from y to x, be designated as respectively (e, x, y) and (e, y, x).Aforesaid operations is all carried out on every limit for G, and the directed edge that is obtained set is designated as E.Definition capacity function c:E → R on E ⁺Deserve to be called and state the capacity function c that defines on figure G and the limit collection thereof and constituted a s-t network, note do N=(G, s, t, c), s, t are called source point and meeting point.Claim to satisfy the function ψ of following condition: E → R is a stream of network: to arbitrarily (e, x, y) ∈ E (x ≠ y), ψ (e, x, y)=-ψ (e, y, x); To any x ∈ V s, t}, $\mathrm{\ψ}(x,V)=\underset{(e,x,y)\∈E}{\mathrm{\Σ}}\mathrm{\ψ}\left(e\right)=0;$ To any e ∈ E, ψ (e)≤c (e).

At this moment, claim $\mathrm{\ψ}(x,V)=\underset{(e,s,x)\∈E}{\mathrm{\Σ}}\mathrm{\ψ}\left(e\right)$ Flow for stream ψ.If $S\⊆V$ And s ∈ S, t ∈ V S, claim S, V the set on limit between the S be that of network N cuts, each bar edge capacity sum during the capacity that cuts is defined as and cuts claims the minimal cut for network N of cutting of capacity minimum in the network N.

Usually, the node of pixel map network in the image, the capacity on difference between the pixel characteristic or the similarity corresponding sides, by with upper type with image mapped to a network, make the energy function of the corresponding visual problem of capacity that network cuts.

Max-flow (minimal cut) theorem.Ford and Fulkerson have proved the fundamental theorem in the network flow theory---max-flow (minimal cut) theorem, the i.e. equivalence of the max-flow of network and minimal cut.At present, can in polynomial time, find the solution network maximum flow problem.

Calendar year 2001, Boykov and Jolly introduce the image segmentation field with figure hugger opinion first, the semi-automatic partition method that cuts based on figure has been proposed, this method is at first obtained part foreground/background sample point with man-machine interaction mode, determine the pdf model of foreground/background then according to sample point, and then definite energy function, construct the s-t network at last and it is carried out the minimum cost cutting, energy function is minimized, obtain final split image.This method can be obtained globally optimal solution and merge multiple priori, has established the basic framework that carries out image segmentation based on figure hugger opinion.

People such as Rother are based on said method, with mixed model (the Gaussian MixtureModel of Gauss, GMM) replace grey level histogram, replace a least estimated with the iterative algorithm that in GMM parameter learning estimation procedure, can evolve and finish energy minimization, proposed the GrabCut algorithm, this method has obtained man-machine interaction effect more easily and more high-precision segmentation ability.

Above-mentioned two kinds of methods all with entire image as process object, to the s-t network of its constructing stereo formula, though obtained globally optimal solution, what definite cost of pdf model and accuracy thereof were having a strong impact on algorithm cuts apart efficient and effect.

People's based target borders such as Ning Xu are in the hypothesis of its certain neighborhood self-energy minimum, behind artificial given initial profile line, with this initial profile line is the annular section that a fixed width is chosen on the limit, carry out the iteration cutting to the s-t network of this annular section formation level formula and to it, obtain object boundary.

People such as Slabaugh have incorporated this priori of geometric configuration of object boundary when the energy function that construct image is cut apart.At first choose a priori outline line similar to the object boundary geometric configuration, with this outline line is the annular section that a fixed width is obtained in the expansion of middle mind-set both sides, then to the s-t network of this annular section constructing stereo formula and merge the priori of object boundary geometric configuration, network is carried out the minimum cost cutting, the gained cutting curve is adjusted match to satisfy the constraint of priori outline line, and the cutting of mode iteration finally obtains object boundary according to this.

The major defect of above-mentioned image partition method is not take into full account the outline line deformation characteristics, and splitting speed is slow.

Summary of the invention

Technical matters to be solved by this invention is to overcome the shortcoming of above-mentioned image partition method, and a kind of color image fast partition method that cuts based on deformation contour model and figure fast and accurately is provided.

Solving the problems of the technologies described above the technical scheme that is adopted comprises the steps:

1. draw current outline line

Selected polygonal region that comprises target outer contour to be split obtains current outline line in image.

2. create the annular section of self-adaptation width

Current outline line is divided into reaches the objective contour line segment and do not reach the objective contour line segment, reached objective contour line segment 1 pixel that inwardly expands, do not reach objective contour line segment at least 5 pixels that inwardly expand, create the annular section of a self-adaptation width.

3. to self-adaptation width annular section structure s-t network

A node in the meeting point t of the source point s of the corresponding s-t network of the outer boundary of self-adaptation width annular section, the corresponding s-t network of inner boundary, the corresponding s-t network of all the other each pixels connects with 8 neighborhood mode cum rights values between neighborhood of nodes.

4. minimum cost cutting

Use the max-flow algorithm, promptly the minimal cut algorithm carries out the minimum cost cutting to the s-t network, obtains a new current outline line.

5. determine to be cut apart the outer contour of target

To last time cutting the new current outline line repeating step 2～step 4 of gained, new current outline line iterative deformation is cut apart the outer contour of target in reaching image.

6. the quilt in the output image is cut apart target

According to the outer contour of being cut apart target target is split.

Of the present invention reached the objective contour line segment and do not reached the objective contour line segment be calculated as follows:

C in the formula (2), C (1) are empty set, and C (0) is initial current outline line, and C (n) is the outline line that the n time cutting obtains, C ^R(n) be the n time cutting obtain reach objective contour line segment, C ^U(n) be the n time cutting obtain do not reach the objective contour line segment, n is 1～70 integer.

The weights that connect with 8 neighborhood mode cum rights values between neighborhood of nodes of the present invention are calculated as follows:

$w_{\mathrm{ij}}=e^{-\frac{{(I_i-I_j)}^2}{{2\mathrm{\σ}}^2}}$

${\mathrm{\σ}}^2=\frac{1}{\left|E\right|}\underset{(i,j)\∈E}{\mathrm{\Σ}}{(I_i-I_j)}^2$

W in the formula _IjBe node i, the weights between the j, I _i, I _jBe node i, the color value of j respective pixel, σ ²Be variance, E is the limit collection of s-t network.

The present invention is improved to the annular section of self-adaptation width with the annular section of fixed width, proposes a kind of coloured image iteration cutting new method based on plane s-t network, unidirectional expansion, self-adaptation width annular section.This method adopts plane formula s-t network, and is more convenient, quick aspect the structure of network, cutting, particularly evident under iterative manner; The unidirectional expansion of outline line can reduce the overlapping of adjacent annular section effectively; After adopting self-adaptation width annular section, only do not cut, make that cutting apart efficient significantly improves reaching the objective contour line segment.Advantage such as that the present invention has is easy and simple to handle, splitting speed is fast, precision is high, noise immunity is strong can be used for cutting apart of various colours and black white image.

Description of drawings

Fig. 1 is the process flow diagram of the inventive method.

Fig. 2 is the concrete dividing method of a width of cloth doll image.

Fig. 3 is the concrete dividing method of the image of a width of cloth hand.

Fig. 4 is the concrete dividing method of a width of cloth stone image.

Embodiment

The present invention is described in more detail below in conjunction with drawings and Examples, but the invention is not restricted to these embodiment.

Embodiment 1

Adopt Matlab R2008b and Microsoft visual studio 2005 C Plus Plus hybrid programmings, realize method described in the invention.Experiment porch is Windows 2003 Server StandardEdition, CPU:Intel Core 2 Duo E6320, RAM:1G.

As follows with Fig. 2 doll image color image fast partition method step that to be the example introduction cut based on deformation contour model and figure:

1. draw current outline line

In image, draw the current outline line 1 of 6 limit shapes that comprises doll outer contour to be split.Current outline line 1 is seen Fig. 2 (a).Fig. 2 (a) is that size is the image of 398 * 284 pixels.

2. create the annular section of self-adaptation width

Current outline line is calculated by following formula, is divided into and reaches objective contour line segment 3 and do not reach objective contour line segment 4:

C in the formula (2), C (1) are empty set, and C (0) is initial current outline line, and C (n) is the outline line that the n time cutting obtains, C ^R(n) be the n time cutting obtain reach objective contour line segment, C ^U(n) be the n time cutting obtain do not reach the objective contour line segment.In the present embodiment, n is 1～10 integer.

To reaching inwardly 1 pixel of expansion of objective contour line segment 3, do not reach inwardly 25 pixels of expansion of objective contour line segment 4, create the annular section 5 of a self-adaptation width, see Fig. 2 (b).Fig. 2 (b) is the annular section 5 of the self-adaptation width of the 3rd cutting gained outline line and formation thereof.

3. to self-adaptation width annular section structure s-t network

A node in the meeting point t of the source point s of the corresponding s-t network of the outer boundary of self-adaptation width annular section 5, the corresponding s-t network of inner boundary, the corresponding s-t network of all the other each pixels, the weights that connect with 8 neighborhood mode cum rights values between neighborhood of nodes are calculated as follows:

$w_{\mathrm{ij}}=e^{-\frac{{(I_i-I_j)}^2}{{2\mathrm{\σ}}^2}}$

${\mathrm{\σ}}^2=\frac{1}{\left|E\right|}\underset{(i,j)\∈E}{\mathrm{\Σ}}{(I_i-I_j)}^2$

W in the formula _IjBe node i, the weights between the j, I _i, I _jBe node i, the color value of j respective pixel, σ ²Be variance, E is the limit collection of s-t network.

4. minimum cost cutting

With max-flow algorithm (being the minimal cut algorithm) the s-t network is carried out the minimum cost cutting, obtain a new current outline line 2.The max-flow algorithm calculates according to the max-flow algorithm in computer research and the development (the 9th phase of September in 2003 " network maximum flow problem progress ").

5. determine to be cut apart the outer contour of target

To last time cutting the new current outline line repeating step 2～step 4 of gained, new current outline line iterative deformation is shunk, and is cut apart the outer contour 6 of doll in reaching image.In the present embodiment, the iteration cutting is 10 times altogether, 1.72 seconds time.Fig. 2 (a) has shown the iterative deformation process of current outline line 1, and Fig. 2 (c) is a segmentation result.

6. the quilt in the output image is cut apart target

Outer contour 6 according to doll splits doll.

Embodiment 2

As follows with the image (size is 265 * 265 pixels) of Fig. 3 hand color image fast partition method step that to be the example introduction cut based on deformation contour model and figure:

In the current outline line of the drafting step 1 of embodiment 1, in image, draw the current outline line 7 of 9 limit shapes that comprises the outer contour of hand to be split.In the annular section step 2 of creating the self-adaptation width, n is 1～70 integer, to reaching inwardly 1 pixel of expansion of objective contour line segment 8, do not reach inwardly 5 pixels of expansion of objective contour line segment 9, create the annular section 10 of a self-adaptation width, other step in this step is identical with embodiment 1.Definite quilt is cut apart in the outer contour step 5 of target, and the iteration cutting is 70 times altogether, and 2.241 seconds time, other step in this step is identical with embodiment 1.Other step is identical with embodiment 1, extracts the outer contour 11 of being cut apart hand.

Embodiment 3

As follows with Fig. 4 stone image (size is 640 * 480 pixels) color image fast partition method step that to be the example introduction cut based on deformation contour model and figure:

In the current outline line of the drafting step 1 of embodiment 1, in image, draw the current outline line 12 of 8 limit shapes that comprises stone outer contour to be split.In the annular section step 2 of creating the self-adaptation width, n is 1～18 integer, to reaching inwardly 1 pixel of expansion of objective contour line segment 14, do not reach inwardly 35 pixels of expansion of objective contour line segment 13, create the annular section 15 of a self-adaptation width, other step in this step is identical with embodiment 1.Definite quilt is cut apart in the outer contour step 5 of target, and the iteration cutting is 18 times altogether, and 2.542 seconds time, other step in this step is identical with embodiment 1.Other step is identical with embodiment 1, extracts the outer contour 16 of being cut apart stone.

According to said method, also can cut apart the arbitrary target of being asked in other colour or the gray level image, but all within protection scope of the present invention.

Claims (3)

1. a color image fast partition method that cuts based on deformation contour model and figure is characterized in that it comprises the steps:

(1) draws current outline line

Selected polygonal region that comprises target outer contour to be split obtains current outline line in image;

(2) annular section of establishment self-adaptation width

Current outline line is divided into reaches the objective contour line segment and do not reach the objective contour line segment, reached objective contour line segment 1 pixel that inwardly expands, do not reach objective contour line segment at least 5 pixels that inwardly expand, create the annular section of a self-adaptation width;

(3) to self-adaptation width annular section structure s-t network

A node in the meeting point t of the source point s of the corresponding s-t network of the outer boundary of self-adaptation width annular section, the corresponding s-t network of inner boundary, the corresponding s-t network of all the other each pixels connects with 8 neighborhood mode cum rights values between neighborhood of nodes;

(4) minimum cost cutting

Use the max-flow algorithm, promptly the minimal cut algorithm carries out the minimum cost cutting to the s-t network, obtains a new current outline line;

(5) definite outer contour of being cut apart target

To last time cutting the new current outline line repeating step of gained (2)～step (4), new current outline line iterative deformation is cut apart the outer contour of target in reaching image;

(6) quilt in the output image is cut apart target

According to the outer contour of being cut apart target target is split.

2. according to the described color image fast partition method that cuts based on deformation contour model and figure of claim 1, it is characterized in that said reached the objective contour line segment and do not reached the objective contour line segment be calculated as follows:

C in the formula (2), C (1) are empty set, and C (0) is initial current outline line, and C (n) is the outline line that the n time cutting obtains, C ^R(n) be the n time cutting obtain reach objective contour line segment, C ^U(n) be the n time cutting obtain do not reach the objective contour line segment, n is 1～70 integer.

3. according to the described color image fast partition method that cuts based on deformation contour model and figure of claim 1, it is characterized in that the weights that connect with 8 neighborhood mode cum rights values between said neighborhood of nodes are calculated as follows:

$w_{\mathrm{ij}}=e^{-\frac{{(I_i-I_j)}^2}{{2\mathrm{\σ}}^2}}$

${\mathrm{\σ}}^2=\frac{1}{\left|E\right|}\underset{(i,j)\∈E}{\mathrm{\Σ}}{(I_i-I_j)}^2$

W in the formula _IjBe node i, the weights between the j, I _i, I _jBe node i, the color value of j respective pixel, σ ²Be variance, E is the limit collection of s-t network.

Images