CN103955945A - Self-adaption color image segmentation method based on binocular parallax and movable outline - Google Patents

Abstract

The invention relates to a self-adaption color image segmentation method based on the binocular parallax and a movable outline, and belongs to the technical field of computer visual processing. The self-adaption color image segmentation method solves the problem that an existing movable outline model is limited for segmenting monocular gray level images and can not be well applied to binocular color images, and the problem that due to the fact that an initial outline is determined mainly relying on prior experience, the initial outline can not be accurately set in a self-adaption mode and the segmentation result is influenced. The self-adaption color image segmentation method based on the binocular parallax and the movable outline is achieved according to the following steps of setting the initial outline on the basis of the binocular parallax in a self-adaption mode, converting the color space, setting up an energy functional based on an improved LCV model, evolving an outline curve, and outputting the segmentation result. The self-adaption color image segmentation method is suitable for three-dimensional image segmentation or three-dimensional video compression preprocessing or target recognition or the like.

Description

Self-adaption colorful image partition method based on binocular parallax and active contour

Technical field

The present invention relates to stereo-picture processing, is a kind of self-adaption colorful image partition method based on binocular parallax and active contour, belongs to computer vision processing technology field.

Background technology

The information major part that people obtain is from vision system, and still, the things that human eye is seen has stereoscopic sensation, and common image is two-dimentional.Along with scientific and technological progress, binocular stereo image occupies a tiny space gradually in people's life, in the application of every field, also becomes more and more important.As object tracing, self-navigation, medical science auxiliary diagnosis, virtual reality, mapping etc.Image Engineering can be divided into three levels, image processing, graphical analysis and image understanding conventionally.And image is cut apart as the committed step of processing image analysis process from image, be also focus and the difficult point of research for a long time always.In recent years, movable contour model, because it possesses the modeling of being easy to and the advantage such as Mathematical is efficient, becomes a large focus of cutting apart field.These class methods are used the priori such as constraint information and position, size and the shape of target obtaining from view data, and are unified in a characteristic extraction procedure, can effectively cut apart target.

According to the difference of contour curve type, movable contour model can be divided into two classes: parametric active contour model and geometric active contour model.Parametric active contour model claims again Snake model, and this class model, to initial position sensitivity, need be arranged near interesting target, and it is also poor in curve evolvement process, to tackle the ability of change in topology.The closed curve that geometric active contour model is plane by the zero level set representations of higher-dimension toroidal function (level set function), adopts the form of level set to describe the evolution of curve, thereby is implied with the ability of change in topology.But only adopt the level set algorithm of marginal information, more responsive to weak edge and discontinuous edge.Chan and Vese are at " IEEE Transactions on Image Processing " 2001,10 (2), the article " Active contours without edges " of delivering on pp.266-277 has proposed a kind of Chan-Vese model, be called for short CV model, utilize the information of homogeneous region similarity, take homogeneous global statistics hypothesis, can be partitioned into preferably the target of weak edge or discontinuous edge.But for the image of nonuniformity, can not obtain correct segmentation result, and can only carry out gray scale and cut apart.On the basis of CV model, Lankton and Tannenbaum are at " IEEE Transactions on Image Processing " 2008,17 (11), the article of delivering on pp.2029-2039 " Localizing region-based active contours " a kind of geometric active contour model based on local message proposed, be called for short LCV model, this model is directly added up local pixel grey scale average, has well solved the problem of cutting apart nonuniformity image.But several movable contour models that proposed are all only confined to cut apart monocular gray level image, can not well be applied to binocular coloured image.

Movable contour model is to be developed and carried out segmentation object by contour curve, and when the initial profile arranging more approaches position, the size and shape of interesting target, its segmentation result precision is higher.And determine when initial profile at present and mainly rely on priori experience, can accurately not realize self-adaptation initial profile position is set, thereby affect segmentation result.

Summary of the invention

The object of the invention is to propose a kind of self-adaption colorful image partition method based on binocular parallax and active contour, to be all only confined to cut apart monocular gray level image for existing movable contour model, can not well be applied to binocular coloured image; While determining initial profile, mainly rely on priori experience, can accurately not realize self-adaptation initial profile position is set, thereby affect the problem of segmentation result.

For solving the problems of the technologies described above adopted technical scheme be:

Self-adaption colorful image partition method based on binocular parallax and active contour of the present invention, realize according to following steps:

Step 1, self-adaptive initial profile based on binocular parallax arrange;

The conversion of step 2, color space;

Step 3, the energy functional of foundation based on improved LCV model;

The evolution of step 4, contour curve;

The output of step 5, segmentation result.

The invention has the beneficial effects as follows:

One, the present invention, to the geometric active contour model based on local message, is called for short LCV model and improves, and is expanded to binocular coloured image.

Two, the LCV model after improvement is introduced binocular parallax in the process that initial profile is set, can initial profile be set correct self-adaptation, compare with the initial profile that relies on prior imformation to obtain, this method can be approached the initial profile of interesting target position, size and shape more, improves the precision of cutting apart binocular image.

Three, the energy that the present invention adds initial profile information to LCV model is general middle as contour shape bound term, the level set function that refers to LCV model is determined by initial profile curve, just can effectively the shape information of target be added in parted pattern with this, improve the efficiency of cutting apart.

Four, the present invention has carried out quantitatively evaluating to efficiency.The in the situation that of identical step-length, it is as shown in table 1 that different models is cut apart the needed iterations of same piece image.As seen from Table 1, the present invention greatly reduces than LCV and two kinds of model iterationses of CV, and evolution speed is very fast.

Five, describedly change by color space, the colouring information that the present invention is utilized is more even, refer to binocular stereo image is transformed in YCbCr color space by rgb color space, brightness in image and colourity are separated, colouring information distributes more even, replace gray average with colourity average, both made full use of color of image information, LCV model is cut apart and is generalized to Color Segmentation by gray scale, retain again former LCV model and can cut apart the advantage of nonuniformity image, obtained gratifying segmentation effect.

Brief description of the drawings

Fig. 1 is process flow diagram of the present invention; Fig. 2 is the segmentation result comparison that the LCV model of different initial profiles is set, and wherein, (a)-left image in (f) is the initial profile arranging, and right image is the segmentation result of LCV model; Fig. 3 is the self-adaptive initial profile method to set up design sketch based on binocular parallax, wherein, (a) 5 pictures of row are input picture, (b) classify the interesting target subject area corresponding to input picture as, (c) classify the initial profile that the present invention arranges input picture as; Fig. 4 is segmentation result comparison of the present invention in YCbCr color space and rgb color space, wherein, (a) 5 pictures of row are input picture, (b) classify the segmentation result of the present invention in YCbCr color space as, (c) classify the segmentation result of the present invention in rgb color space as; Fig. 5 is that the present invention and LCV model and CV model are cut apart Venus image, wherein, (a) row four width images are respectively input picture from left to right, the initial profile that the present invention arranges, the initial profile that the initial profile that LCV model arranges and CV model arrange, (b) row four width images are respectively true picture from left to right, the segmentation result that the present invention obtains, the segmentation result that the segmentation result that LCV model obtains and CV model obtain; Fig. 6 is that the present invention and LCV model and CV model are cut apart Tsukuba-lamp image, wherein, (a) row four width images are respectively input picture from left to right, the initial profile that the present invention arranges, the initial profile that the initial profile that LCV model arranges and CV model arrange, (b) row four width images are respectively true picture from left to right, the segmentation result that the present invention obtains, the segmentation result that the segmentation result that LCV model obtains and CV model obtain; Fig. 7 is that the present invention and LCV model and CV model are cut apart Tsukuba-statue image, wherein, (a) row four width images are respectively input picture from left to right, the initial profile that the present invention arranges, the initial profile that the initial profile that LCV model arranges and CV model arrange, (b) row four width images are respectively true picture from left to right, the segmentation result that the present invention obtains, the segmentation result that the segmentation result that LCV model obtains and CV model obtain; Fig. 8 is that the present invention and LCV model and CV model are cut apart Sawtooth image; Fig. 9 is that the present invention and LCV model and CV model are cut apart Poster image, wherein, (a) row four width images are respectively input picture from left to right, the initial profile that the present invention arranges, the initial profile that the initial profile that LCV model arranges and CV model arrange, (b) row four width images are respectively true picture from left to right, the segmentation result that the present invention obtains, the segmentation result that the segmentation result that LCV model obtains and CV model obtain; Figure 10 is CAR and the BAR value comparison diagram that the present invention and LCV model and CV model obtain for different images, wherein, horizontal ordinate is different corresponding CAR and the BAR of input picture, image name is respectively from left to right: Venus, Tsukuba-lamp, Tsukuba-statue, Sawtooth, Poster, ordinate is CAR and BAR value.

Embodiment

Embodiment one: the self-adaption colorful image partition method based on binocular parallax and active contour described in present embodiment, realize according to following steps:

Step 1, self-adaptive initial profile based on binocular parallax arrange;

The conversion of step 2, color space;

Step 3, the energy functional of foundation based on improved LCV model;

The evolution of step 4, contour curve;

The output of step 5, segmentation result.Understand present embodiment in conjunction with Fig. 1.

Embodiment two: present embodiment is different from embodiment one: the initial profile setting described in step 1, realizes according to following steps:

Step 1 (one), look like as target image taking left view in binocular stereo image, it is reference picture that right view looks like, and adopts adaptive weighted Stereo Matching Algorithm to obtain the disparity map of left view picture in binocular stereo image;

Step 1 (two), disparity map is carried out to Threshold segmentation, extract interested destination object region, then utilize medium filtering to suppress the noise in disparity map;

Step 1 (three), the interesting target zone boundary obtaining are set to the initial profile of movable contour model, detailed process is: select target subject area, body surface is all generally smooth, therefore on body surface, the projection of each point on image is continuous, and its parallax is also continuous; According to this disparity continuity constraint condition, can be to extracting respectively in the destination object of different parallax planes in disparity map; In disparity map, choose a region n ₁× n ₂, calculate the average of parallax value in this region z _ifor chosen area n in disparity map ₁× n ₂interior any pixel, the wherein position of i represent pixel point, as z ₁represent first pixel, z ₅represent the 5th pixel; Destination object region is adjudicated by formula (1):

$\mathrm{object}\left(z_i\right)=\left\{\begin{array}{cc}1,& |d\left(z_i\right)-\stackrel{\‾}{d}|\≤\mathrm{\δ}\\ 0,& |d\left(z_i\right)-\stackrel{\‾}{d}|>\mathrm{\δ}\end{array}\right.---\left(1\right)$

Wherein, d (z _i) be pixel z _iparallax value, δ be set threshold value, object (z _i) value is 1 region is destination object region, object (z _i) value is 0 region is background area; Adopt morphological method to be further processed destination object region, obtain one and comprise the comparatively smooth irregular area of destination object full detail and edge,, the boundary profile that finally obtains destination object with erosion operation is set to the initial profile C of movable contour model, i.e. closed contour curve C.Understand present embodiment in conjunction with Fig. 2, Fig. 3, other step and parameter are identical with embodiment one.

Embodiment three: present embodiment is different from embodiment one or two: the conversion of the color space described in step 2, is realized by following rgb color space and YCbCr color space conversion formula:

Y＝0.299R+0.587G+0.114B

Cb＝0.564(B-Y)

Cr=0.713 (R-Y), wherein, Y, Cb, Cr represents respectively the brightness of YCbCr color space, three components of chroma blue and red color; R, G, B represents respectively the redness of rgb color space, green, blue three components.Understand present embodiment in conjunction with Fig. 4, other step and parameter are identical with embodiment one or two.

Embodiment four: present embodiment is different from one of embodiment one to three: the energy functional of the foundation described in step 3 based on improving LCV model, realize according to following steps:

Step 3 (one), LCV model are the movable contour model based on local, obtain the tool of the regional area of image I

Body process is:

Utilize the regional area of fundamental function B (x, y) acquisition image I,

In formula (2), definition B (x, y) is a ball, to image I mask, obtains the regional area of an image I; X represents the central point of this bulbous region, and y represents another spatial point, and r represents the radius of ball;

In the time that the distance of x and y is less than radius r, the value of fundamental function B (x, y) is 1, and representation space point y is in the inside of bulbous region; Otherwise be 0, representation space point does not belong to bulbous region;

The energy functional mathematic(al) representation of step 3 (two), definition LCV model is as follows:

$E_{\mathrm{LCV}}\left(\mathrm{\φ}\right)={\∫}_{{\mathrm{\Ω}}_x}\mathrm{\δ\φ}\left(x\right){\∫}_{{\mathrm{\Ω}}_y}B(x,y)\·F(I\left(y\right),\mathrm{\φ}\left(y\right))\mathrm{dydx}+\mathrm{\λ}{\∫}_{{\mathrm{\Ω}}_x}\mathrm{\δ\φ}\left(x\right)\left|\right|\▿\mathrm{\φ}\left(x\right)\left|\right|\mathrm{dx}---\left(3\right)$

Wherein, Ω representative graph image field, φ (x) represents closed evolution curve for level set function, and δ φ (x) is a level and smooth Dirac function later, can ensure that δ φ (x) numerical value of putting on all non-zero level collection in image area is all approximately zero, the uniqueness of contour curve while guaranteeing curve evolvement, B (x, y) represents the regional area that image I mask is obtained, and F represents a general internal energy function, λ is the weight coefficient of regularization term for the gradient of level set function;

In formula (3), Section 1 for local energy item, internal energy function F is only calculated the regional area of B (x, y) mask, and guiding closed contour curve reduces direction to energy and moves, Section 2 for regularization term, in order to adjust the shape of evolution curve;

It is the left view picture in binocular stereo image that step 3 (three), setting image I belong to image area Ω, be defined as input picture, O is the binary map that the self-adaptive initial profile method to set up based on binocular parallax obtains initial position, closed contour curve C is expressed as zero level collection C={x| φ (the x)=0} of level set function, and input picture I is divided into target area and background area;

The pixel colourity average of objective definition region and background area is respectively c ₁and c ₂:

$c_1=\frac{{\∫}_{{\mathrm{\Ω}}_O}{I}_O(x,y)\·\mathrm{H\φ}\left(x\right)\·\mathrm{dxdy}}{{\∫}_{{\mathrm{\Ω}}_O}\mathrm{H\φ}\left(x\right)\mathrm{dxdy}}---\left(4\right)$

$C_2=\frac{{\∫}_{{\mathrm{\Ω}}_O}{I}_O(x,y)\·(1-\mathrm{H\φ}\left(x\right))\·\mathrm{dxdy}}{{\∫}_{{\mathrm{\Ω}}_O}(1-\mathrm{H\φ}\left(x\right))\mathrm{dxdy}}---\left(5\right)$

Wherein, IO is the regional area that O obtains image I mask, $H\left(\mathrm{\&phi;}\right)=\left\{\begin{array}{c}1,\mathrm{\&phi;}\&GreaterEqual;0\\ 0,\mathrm{\&phi;}<0\end{array}\right.$ It is Heaviside function, divide evolution region, in the time of φ >0, show the outside of closed curve on real profile border, closed curve needs toe-in to arrive real profile border, in the time of φ=0, shows that closed curve overlaps with real profile border, H when both of these case (φ)=1, calculating closed curve inside is the pixel colourity average c of target area ₁; In the time of φ <0, show the inside of closed curve on real profile border, closed curve need to expand outwardly and arrive real profile border, now H (φ)=0, calculating closed curve outside is the pixel colourity average c of background area ₂;

Step 3 (four), based on improved LCV model, set up energy functional mathematic(al) representation and be

$\begin{array}{c}E\left(\mathrm{\&phi;}\right)={\&Integral;}_{{\mathrm{\&Omega;}}_x}\mathrm{\&delta;\&phi;}\left(x\right){\&Integral;}_{{\mathrm{\&Omega;}}_O}{I}_O(x,y)[\mathrm{H\&phi;}\left(y\right){(I_O\left(y\right)-c_1)}^2+(1-\mathrm{H\&phi;}\left(y\right)){(I_O\left(y\right)-c_2)}^2]+\mathrm{\&lambda;}{\&Integral;}_{{\mathrm{\&Omega;}}_x}\mathrm{\&delta;\&phi;}\left(x\right)\left|\right|\&dtri;\mathrm{\&phi;}\left(x\right)\left|\right|\mathrm{dx}\\ +\mathrm{\&mu;}{\&Integral;}_{\mathrm{\&Omega;}}\frac{1}{2}{\left(\right|\&dtri;\mathrm{\&phi;}\left(x\right)|-1)}^2\mathrm{dxdy}\end{array}---\left(6\right)$

Wherein, Ω representative graph image field, φ (x) represents closed evolution curve for level set function, δ φ (x) is a level and smooth Dirac function later, can ensure that δ φ (x) numerical value of putting on all non-zero level collection in image area is all approximately zero, the uniqueness of contour curve while guaranteeing curve evolvement, IO is the region that O obtains image I mask $H\left(\mathrm{\&phi;}\right)=\left\{\begin{array}{c}1,\mathrm{\&phi;}\&GreaterEqual;0\\ 0,\mathrm{\&phi;}<0\end{array}\right.$ Be Heaviside function, divide evolution region, c ₁and c ₂the pixel colourity average that is respectively objective definition region and background area, λ is regularization term weight coefficient, μ is energy penalty term coefficient, for the gradient of level set function; In formula (6), Section 1 ${\&Integral;}_{{\mathrm{\&Omega;}}_x}\mathrm{\&delta;\&phi;}\left(x\right){\&Integral;}_{{\mathrm{\&Omega;}}_O}{I}_O(x,y)[\mathrm{H\&phi;}\left(y\right){(I_O\left(y\right)-c_1)}^2+(1-\mathrm{H\&phi;}\left(y\right)){(I_O\left(y\right)-c_2)}^2]$ For the energy term based on local, guiding closed contour curve reduces direction along energy and moves, Section 2 for regularization term, in order to adjust the shape of evolution curve, keep the smooth of level set; Section 3 for the energy penalty term of introducing, make level set function approximate remaining apart from sign function in evolutionary process, avoid reinitializing.Other step and parameter are identical with one of embodiment one to three.

Embodiment five: present embodiment is different from one of embodiment one to four: the contour curve described in step 4 develops and realizes according to following steps: contour curve represents with level set function in this model, so contour curve evolutionary process is the solution procedure of level set movements equation.Be specially: in the chromatic component of YCbCr space, to being carried out curve evolvement by the regional area of O mask, according to energy functional first variation, the level set movements equation of deriving contour curve evolution is

$\begin{array}{c}\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}\left(x\right)=\mathrm{\&delta;\&phi;}\left(x\right){\&Integral;}_{{\mathrm{\&Omega;}}_O}{I}_O(x,y)\&CenterDot;\mathrm{\&delta;\&phi;}\left(y\right)\&CenterDot;[{(I_O\left(y\right)-c_1)}^2-{(I_O\left(y\right)-c_2)}^2]\mathrm{dy}\\ +\mathrm{\&mu;}|\left(\mathrm{\&delta;\&phi;}\left(x\right)\frac{\&dtri;\mathrm{\&phi;}\left(x\right)}{|\&dtri;\mathrm{\&phi;}\left(x\right)|}\right)+\left({\&dtri;}^2\right)\mathrm{\&phi;}\left(x\right)-\mathrm{div}\frac{\&dtri;\mathrm{\&phi;}\left(x\right)}{|\&dtri;\mathrm{\&phi;}\left(x\right)|}|\end{array}---\left(7\right)$

Wherein, Ω representative graph image field, φ (x) represents closed evolution curve for level set function, δ φ (x) is a level and smooth Dirac function later, can ensure that δ φ (x) numerical value of putting on all non-zero level collection in image area is all approximately zero, the uniqueness of contour curve while guaranteeing curve evolvement, I _ofor the regional area that O obtains image I mask, c ₁and c ₂the pixel colourity average that is respectively objective definition region and background area, μ is penalty term coefficient, for Laplace operator, for the bent curvature of a curve that develops.Other step and parameter are identical with one of embodiment one to four.

Embodiment six: present embodiment is different from one of embodiment one to five: the segmentation result output described in step 5 realizes according to following steps: calculating formula (7), express the solution of partial differential equation with level set function φ (x), it is a closed curve being represented by level set function φ (x) that equation unique solution is presented in image, image is divided into destination object region and background area by this closed curve, and wherein destination object region is exactly the image of cutting apart finally obtaining.Other step and parameter are identical with one of embodiment one to five.

Fig. 5-Fig. 9 is the comparison of model of the present invention and LCV model and CV model segmentation result, in order to verify validity of the present invention, adopt CAR, be public domain rate (Common Area Rate) and BAR, background rate (Background Area Rate) is assessed the precision of the segmentation result that three kinds of models obtain.Public domain rate formula and background rate formula are respectively

CAR＝(S _con∩S _truth)/S _truth (1)

BAR＝[S _con-(S _con∩S _truth)]/S _con (2)

Wherein, S _conthe area of the objective contour interior zone obtaining for different model silhouette curve evolvements, S _truthfor the area of real goal profile interior zone.In formula (1) and formula (2), public domain rate CAR and background rate BAR are all the numerical value of a 0-1, and CAR is larger, and the destination object that representative model segmentation result comprises is more, and BAR is less, and the background that representative model segmentation result comprises is fewer.Be that the larger BAR of CAR is less, the segmentation precision of model is higher.

Figure 10 is to be the CAR that obtains for different images of model of the present invention and LCV model and CV model and the comparison of BAR.Basic identical compared with two kinds of algorithms of the destination object that the segmentation result that the present invention obtains comprises and other, but the background comprising greatly reduces.Therefore,, for binocular stereo image, the model that the present invention proposes is significantly increased compared with conventional model is on segmentation precision.

Table 1 the present invention and other models are cut apart speed ratio

。

Claims (6)

1. the self-adaption colorful image partition method based on binocular parallax and active contour, is characterized in that described method realizes according to following steps:

Step 1, self-adaptive initial profile based on binocular parallax arrange;

The conversion of step 2, color space;

Step 3, the energy functional of foundation based on improved LCV model;

The evolution of step 4, contour curve;

The output of step 5, segmentation result.

2. the self-adaption colorful image partition method based on binocular parallax and active contour according to claim 1, is characterized in that the initial profile setting described in step 1, realizes according to following steps:

Step 1 (one), look like as target image taking left view in binocular stereo image, it is reference picture that right view looks like, and adopts adaptive weighted Stereo Matching Algorithm to obtain the disparity map of left view picture in binocular stereo image;

Step 1 (two), disparity map is carried out to Threshold segmentation, extract interested destination object region, then utilize medium filtering to suppress the noise in disparity map;

Step 1 (three), the interesting target zone boundary obtaining are set to the initial profile of movable contour model; Detailed process is: select target subject area, to extracting respectively in the destination object of different parallax planes in disparity map; In disparity map, choose a region n ₁× n ₂, calculate the average of parallax value in this region z _ifor chosen area n in disparity map ₁× n ₂in any pixel, the wherein position of i represent pixel point, destination object region is adjudicated by formula (1):

$\mathrm{object}\left(z_i\right)=\left\{\begin{array}{cc}1,& |d\left(z_i\right)-\stackrel{\&OverBar;}{d}|\&le;\mathrm{\&delta;}\\ 0,& |d\left(z_i\right)-\stackrel{\&OverBar;}{d}|>\mathrm{\&delta;}\end{array}\right.---\left(1\right)$

Wherein, d (z _i) be pixel z _iparallax value, δ be set threshold value, object (z _i) value is 1 region is destination object region, object (z _i) value is 0 region is background area; Adopt morphological method to be further processed destination object region, obtain one and comprise the comparatively smooth irregular area of destination object full detail and edge, the boundary profile that finally obtains destination object region with erosion operation is set to the initial profile C of movable contour model, i.e. closed contour curve C.

3. the self-adaption colorful image partition method based on binocular parallax and active contour according to claim 2, is characterized in that the conversion of the color space described in step 2, is realized by following rgb color space and YCbCr color space conversion formula:

Y＝0.299R+0.587G+0.114B

Cb＝0.564(B-Y)

Cr=0.713 (R-Y), wherein, Y, Cb, Cr represents respectively the brightness of YCbCr color space, three components of chroma blue and red color; R, G, B represents respectively the redness of rgb color space, green, blue three components.

4. the self-adaption colorful image partition method based on binocular parallax and active contour according to claim 3, is characterized in that, the energy functional of the foundation described in step 3 based on improved LCV model realized according to following steps:

Step 3 (one), the detailed process of obtaining the regional area of image I are:

Utilize the regional area of fundamental function B (x, y) acquisition image I,

In formula (2), definition B (x, y) is a ball, to image I mask, obtains the regional area of an image I; X represents the central point of this bulbous region, and y represents another spatial point, and r represents the radius of ball;

In the time that the distance of x and y is less than radius r, the value of fundamental function B (x, y) is 1, and representation space point y is in the inside of bulbous region; Otherwise be 0, representation space point does not belong to bulbous region;

The mathematic(al) representation of step 3 (two), definition LCV model energy functional is as follows:

$E_{\mathrm{LCV}}\left(\mathrm{\&phi;}\right)={\&Integral;}_{{\mathrm{\&Omega;}}_x}\mathrm{\&delta;\&phi;}\left(x\right){\&Integral;}_{{\mathrm{\&Omega;}}_y}B(x,y)\&CenterDot;F(I\left(y\right),\mathrm{\&phi;}\left(y\right))\mathrm{dydx}+\mathrm{\&lambda;}{\&Integral;}_{{\mathrm{\&Omega;}}_x}\mathrm{\&delta;\&phi;}\left(x\right)\left|\right|\&dtri;\mathrm{\&phi;}\left(x\right)\left|\right|\mathrm{dx}---\left(3\right)$

Wherein, Ω representative graph image field, φ (x) represents closed evolution curve for level set function, and δ φ (x) is a level and smooth Dirac function later, can ensure that δ φ (x) numerical value of putting on all non-zero level collection in image area is all approximately zero, the uniqueness of contour curve while guaranteeing curve evolvement, B (x, y) represents the regional area that image I mask is obtained, and F represents a general internal energy function, λ is the weight coefficient of regularization term for the gradient of level set function;

In formula (3), Section 1 for local energy item, internal energy function F is only calculated the regional area of B (x, y) mask, and guiding closed contour curve reduces direction to energy and moves, Section 2 for regularization term, in order to adjust the shape of evolution curve;

It is the left view picture in binocular stereo image that step 3 (three), setting image I belong to image area Ω, be defined as input picture, O is the binary map that the self-adaptive initial profile method to set up based on binocular parallax obtains initial position, closed contour curve C is expressed as zero level collection C={x| φ (the x)=0} of level set function, and input picture I is divided into target area and background area;

The pixel colourity average of objective definition region and background area is respectively c ₁and c ₂:

$c_1=\frac{{\&Integral;}_{{\mathrm{\&Omega;}}_O}{I}_O(x,y)\&CenterDot;\mathrm{H\&phi;}\left(x\right)\&CenterDot;\mathrm{dxdy}}{{\&Integral;}_{{\mathrm{\&Omega;}}_O}\mathrm{H\&phi;}\left(x\right)\mathrm{dxdy}}---\left(4\right)$

$C_2=\frac{{\&Integral;}_{{\mathrm{\&Omega;}}_O}{I}_O(x,y)\&CenterDot;(1-\mathrm{H\&phi;}\left(x\right))\&CenterDot;\mathrm{dxdy}}{{\&Integral;}_{{\mathrm{\&Omega;}}_O}(1-\mathrm{H\&phi;}\left(x\right))\mathrm{dxdy}}---\left(5\right)$

Wherein, I _oregional area image I mask being obtained for O, $H\left(\mathrm{\&phi;}\right)=\left\{\begin{array}{c}1,\mathrm{\&phi;}\&GreaterEqual;0\\ 0,\mathrm{\&phi;}<0\end{array}\right.$ It is Heaviside function, divide evolution region, in the time of φ >0, show the outside of closed curve on real profile border, closed curve needs toe-in to arrive real profile border, in the time of φ=0, show that closed curve overlaps with real profile border, when both of these case, H (φ)=1, calculating closed curve inside is the pixel colourity average c of target area ₁; In the time of φ <0, show the inside of closed curve on real profile border, closed curve need to expand outwardly and arrive real profile border, now H (φ)=0, calculating closed curve outside is the pixel colourity average c of background area ₂; Step 3 (four), based on improved LCV model, set up energy functional mathematic(al) representation and be

$\begin{array}{c}E\left(\mathrm{\&phi;}\right)={\&Integral;}_{{\mathrm{\&Omega;}}_x}\mathrm{\&delta;\&phi;}\left(x\right){\&Integral;}_{{\mathrm{\&Omega;}}_O}{I}_O(x,y)[\mathrm{H\&phi;}\left(y\right){(I_O\left(y\right)-c_1)}^2+(1-\mathrm{H\&phi;}\left(y\right)){(I_O\left(y\right)-c_2)}^2]+\mathrm{\&lambda;}{\&Integral;}_{{\mathrm{\&Omega;}}_x}\mathrm{\&delta;\&phi;}\left(x\right)\left|\right|\&dtri;\mathrm{\&phi;}\left(x\right)\left|\right|\mathrm{dx}\\ +\mathrm{\&mu;}{\&Integral;}_{\mathrm{\&Omega;}}\frac{1}{2}{\left(\right|\&dtri;\mathrm{\&phi;}\left(x\right)|-1)}^2\mathrm{dxdy}\end{array}---\left(6\right)$

Wherein, Ω representative graph image field, φ (x) represents closed evolution curve for level set function, δ φ (x) is a level and smooth Dirac function later, I _oregion image I mask being obtained for O, $H\left(\mathrm{\&phi;}\right)=\left\{\begin{array}{c}1,\mathrm{\&phi;}\&GreaterEqual;0\\ 0,\mathrm{\&phi;}<0\end{array}\right.$ Be Heaviside function, divide evolution region, c ₁and c ₂the pixel colourity average that is respectively objective definition region and background area, λ is regularization term weight coefficient, μ is energy penalty term coefficient, for the gradient of level set function; In formula (6), Section 1

${\&Integral;}_{{\mathrm{\&Omega;}}_x}\mathrm{\&delta;\&phi;}\left(x\right){\&Integral;}_{{\mathrm{\&Omega;}}_O}{I}_O(x,y)[\mathrm{H\&phi;}\left(y\right){(I_O\left(y\right)-c_1)}^2+(1-\mathrm{H\&phi;}\left(y\right)){(I_O\left(y\right)-c_2)}^2]$ For the energy term based on local, guiding closed contour curve reduces direction along energy and moves, Section 2 for regularization term, in order to adjust the shape of evolution curve, keep the smooth of level set; Section 3 for the energy penalty term of introducing, make level set function approximate remaining apart from sign function in evolutionary process, avoid reinitializing.

5. the self-adaption colorful image partition method based on binocular parallax and active contour according to claim 4, it is characterized in that contour curve described in step 4 develops realizes according to following steps: contour curve represents with level set function in this model, be specially: in the chromatic component of YCbCr space to being carried out curve evolvement by the regional area of O mask, according to energy functional first variation, the level set movements equation of deriving contour curve evolution is

$\begin{array}{c}\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}\left(x\right)=\mathrm{\&delta;\&phi;}\left(x\right){\&Integral;}_{{\mathrm{\&Omega;}}_O}{I}_O(x,y)\&CenterDot;\mathrm{\&delta;\&phi;}\left(y\right)\&CenterDot;[{(I_O\left(y\right)-c_1)}^2-{(I_O\left(y\right)-c_2)}^2]\mathrm{dy}\\ +\mathrm{\&mu;}|\left(\mathrm{\&delta;\&phi;}\left(x\right)\frac{\&dtri;\mathrm{\&phi;}\left(x\right)}{|\&dtri;\mathrm{\&phi;}\left(x\right)|}\right)+\left({\&dtri;}^2\right)\mathrm{\&phi;}\left(x\right)-\mathrm{div}\frac{\&dtri;\mathrm{\&phi;}\left(x\right)}{|\&dtri;\mathrm{\&phi;}\left(x\right)|}|\end{array}---\left(7\right)$

Wherein, Ω representative graph image field, φ (x) represents closed evolution curve for level set function, δ φ (x) is a level and smooth Dirac function later, can ensure that δ φ (x) numerical value of putting on all non-zero level collection in image area is all approximately zero, the uniqueness of contour curve while guaranteeing curve evolvement, I _ofor the regional area that O obtains image I mask, c ₁and c ₂the pixel colourity average that is respectively objective definition region and background area, μ is penalty term coefficient, for Laplace operator, for the bent curvature of a curve that develops.

6. the self-adaption colorful image partition method based on binocular parallax and active contour according to claim 5, it is characterized in that the segmentation result output described in step 5 realizes according to following steps: calculating formula (7), express the solution of partial differential equation with level set function φ (x), it is a closed curve being represented by level set function φ (x) that equation unique solution is presented in image, image is divided into destination object region and background area by this closed curve, and wherein destination object region is exactly the image of cutting apart finally obtaining.