CN101447076B - Method for partitioning interested areas in WEB image - Google Patents

Abstract

The invention discloses a method for portioning interested areas in WEB image, comprising the steps as follows: a novel deformation model is established so as to solve the problems of sunken areas and pseudo-boundary and improve the convergence precision of active contour; secondly, a clustering method in a colorful space Lab on the basis of image sub-block, color and vein characteristic is provided so as to realize the coarse partition to the interested areas and automatic gaining of initial position of the active contour, to accelerate the execution speed of the deformation model, and to avoid the interference of noise and isolated edges at the same time; finally, the method provides two execution strategies of the deformation model, which effectively partitions the complex and irregular areas in WEB images and leads the deformation model to have more flexibility and practicability. The method sufficiently combines the advantages of deformation model based on the boundary and the clustering based on the areas, effectively improves the partition precision of the interested areas in images and reduces the whole partition execution time.

Description

The dividing method of area-of-interest in a kind of WEB image

Technical field

The present invention relates to image processing techniques, especially relate to the dividing method of area-of-interest in a kind of WEB image.

Background technology

Flourish along with internet and social informatization all has the digital picture resource of magnanimity to produce from various image capturing systems every day, and (WEB is the abbreviation of World Wide Web, just WWW how to effectively utilize these WEB images; The WEB image is meant the image that is present on the WWW) be not only the most urgent demand of user, and be the top priority that present image is understood.Under considerable situation, the user also is indifferent to the expressed whole meaning of image, and more attention is the area-of-interest that has specific meanings in the image.Image partition method commonly used is broadly divided into two classes: based on cutting apart of zone and cutting apart based on the border.Although the partitioning algorithm of many maturations is arranged at present, limitation and the specific aim of himself arranged all.

Image segmentation based on deformation model is a kind of dividing method based on the border, what adopt is top-down target extractive technique, characteristics of image, object outline and priori can be melted in the unified cutting procedure, be suitable for discerning the area-of-interest of any boundary shape.From people such as Kass since the deformation model (traditional deformation model) that proposes first in the article of on " Journalof Computer Vision ", delivering in 1987 " Snake:Active Contour Models " based on the active contour framework, since its have allow directly mutual with the user, be easy to add shape constraining, compactness and other image partition method such as curve representation form is incomparable flexibly, become a focus of recent this area research.Yet traditional deformation model also has the defective of himself, and for example: the reference position to active contour is comparatively responsive; Can't converge to the border of sunk area; Be subjected to the influence of Local Extremum easily.

In order to overcome the above problems, researchers have proposed some improved models, wherein the most representative have following three kinds: people such as Cohen are by introducing moving of pressure (balloon power) the control active contour that outwards expands or inwardly shrink, and make every effort to overcome the influence of having obeyed the Local Extremum that isolated limit or noise cause by means of balloon.It is a kind of apart from potential energy power that Cohen and Cohen proposes, and the distance map figure that wherein is used for defining potential energy is obtained by each pixel of image and the distance calculation at reference mark on its nearest active contour.But these two kinds of methods are when cutting apart the area-of-interest with depression border, and model can produce power opposite on the horizontal direction when the approaching depression of active contour, makes active contour can't correctly converge to the border of area-of-interest.People such as Xu propose to enlarge with gradient vector flow (GVF) catching range of traditional deformation model, efficiently solve the dependence of deformation model to the active contour initial position, but the diffusion process speed of gradient vector is slow among the GVF, also is difficult to approach the border of dark sunk area.Secondly, these methods all are to cut apart around gray level image, and when needs were cut apart coloured image, method commonly used was exactly to transfer gray level image earlier to.Yet, the WEB image has the diversity of color, the changeableness of zone-texture and the characteristics such as scrambling of object shapes, is subjected to factor such as noise when adding fuzzy and the inhomogeneous and imaging of image itself and makes all be difficult to obtain desirable segmentation effect on these deformation models.Therefore, accurately be partitioned into the area-of-interest in the WEB image, need the bottom perceptual property and the high-layer semantic information of abundant combining target, seek the dividing method that a kind of and human perception is more pressed close to.

Summary of the invention

Because area-of-interest has abundant color and characteristics such as texture, irregular shape usually in the WEB image, cut apart the effect that can't obtain based on what existing deformation model carried out.The object of the present invention is to provide the dividing method of area-of-interest in a kind of WEB image, utilized color, texture and the style characteristic of area-of-interest in the WEB image, remedied traditional deformation model and can't correctly cut apart the depression border, the initial position dependence of active contour is subjected to defectives such as Local Extremum interference greatly, easily; Also improved simultaneously the deficiency that the gray scale split plot design is cut apart coloured image.

The technical solution used in the present invention is:

The dividing method of area-of-interest in a kind of WEB image comprises following content:

1) utilizing color gradient definition image energy, is auxilliary structure bound energy based on triangle heart gravitation, active contour center gravitation, sets up the deformation model that meets the WEB picture characteristics;

2) in the Lab of color space, extract color and texture information, utilize automatic cluster to serve as initialization operation, realize coarse segmentation area-of-interest in the image;

3) cluster result is carried out binary conversion treatment, produce the two-value mask figure that only constitutes, adopt the generation rays method on the foreground area border, to choose the reference mark, obtain the initial position of active contour in the deformation model by prospect and background;

4) adopt two sections implementation strategies to implement the concrete application of deformation model, obtain the accurate segmentation result of area-of-interest in the image.

Described foundation meets the deformation model of WEB picture characteristics, setting up in the process of deformation model, and the structure energy function $E_{\mathrm{snake}}={\∫}_0^1{E}_{\mathrm{snake}}\left(v\left(s\right)\right)\mathrm{ds}={\∫}_0^1(E_{\mathrm{int}}\left(v\left(s\right)\right)+E_{\mathrm{image}}\left(v\left(s\right)\right)+E_{\mathrm{con}}\left(v\left(s\right)\right))\mathrm{ds}$ Utilize chrominance information definition image energy $E_{\mathrm{image}}=-|\▿C_H(x,y)|$ Utilize reference mark v _iPoint to triangle Δ v _I-1v _iv _I+1Heart o _iVector definition curvature, obtain the bound energy that produces by triangle heart gravitation $E_{\mathrm{con}}\left(i\right)=\left|\stackrel{\→}{v_i{o}_i}\right|$ , and increase the three point on a straight line judgement, obtain coming from the auxiliary bound energy of active contour center gravitation $E_{\mathrm{con}}\left(i\right)=f_i\cos (\stackrel{\→}{{\mathrm{cv}}_i},\stackrel{\→}{{\mathrm{cv}}_{i+1}})=(f_i/\left(\left|\stackrel{\→}{{\mathrm{cv}}_i}\right|\left|\stackrel{\→}{{\mathrm{cv}}_{i+1}}\right|\right))(\stackrel{\→}{{\mathrm{cv}}_i}\·\stackrel{\→}{{\mathrm{cv}}_{i+1}}).$

Described realization in the coarse segmentation process of area-of-interest, utilizes color space Lab to extract color characteristic to the coarse segmentation of area-of-interest in the image in i image subblock

With

Utilize the wavelet transformation texture feature extraction

With

, and utilize the k-means algorithm image subblock of M * M to be distinguished in different bunches according to the sextuple feature of prior extraction, bunch number select to adopt adaptive mode.

The described initial position that obtains active contour in the deformation model, in the automatic acquisition process of active contour initial position, cluster result is carried out binary conversion treatment, produce the two-value mask figure that only constitutes by prospect and background, and be starting point with the center of prospect region, outwards generate L bar ray, the angle between adjacent two rays is 2 π/L, and the curve that is formed by connecting by the intersection point on these rays and foreground area border constitutes the initial position of active contour in the deformation model;

Two sections implementation strategies of described employing are implemented the concrete application of deformation model, in the implementation process of two sections implementation strategies of deformation model, do not add bound energy in first section, to reduce the extra cost of calculating curvature, utilize image energy to order about the active contour convergence; Add bound energy in second section, obtaining the global convergence of active contour, and in first section with the bigger quick convergence of reference mark sampling spacing; With the less slow convergence of sampling spacing, area-of-interest accurately cuts apart in the realization image in second section.

The beneficial effect that the present invention has is:

The present invention has made full use of color, texture and the style characteristic of area-of-interest in the WEB image, has remedied traditional deformation model and can't correctly cut apart the depression border, the initial position dependence of active contour is subjected to defectives such as Local Extremum interference greatly, easily; Also improved simultaneously the deficiency that the gray scale split plot design is cut apart coloured image.Utilization is carried out initialization based on the automatic cluster in zone, and area-of-interest in the image is carried out coarse segmentation; Utilization is carried out essence to area-of-interest in the image and is cut apart based on the deformation model generation deformation on border, has effectively improved segmentation precision, has reduced the whole execution time of cutting apart.Be applied at last on composograph and the true picture, obtained gratifying segmentation result.

Description of drawings

Fig. 1 is the process flow diagram of the dividing method of area-of-interest in a kind of WEB image of the present invention.

Fig. 2 is the segmentation result comparisons of various deformation models to the sunk area object.

Fig. 3 is the segmentation result comparisons (275 iteration) of various deformation models to area-of-interest in the image.

Fig. 4 is the deformation model that proposes among the present invention when getting different image subblocks, and relatively (iterations: first section is 100 to the segmentation result of area-of-interest in the image; Second section is 50).

Fig. 5 is the influence of the acquisition methods of different active contour initial position to various deformation models.

Embodiment

Below in conjunction with drawings and Examples the present invention is described in further detail.

As shown in Figure 1, the dividing method of area-of-interest mainly may further comprise the steps in a kind of WEB image of the present invention:

1, the foundation of deformation model:

Utilizing color gradient definition image energy, is auxilliary structure bound energy based on triangle heart gravitation, active contour center gravitation, sets up the deformation model that meets the WEB picture characteristics, specific as follows:

Active contour in the tradition deformation model is a telescopic curve, and by its energy function being minimized control curve deformation process, the closed curve with least energy that obtains at last is exactly the border of area-of-interest.From the mathematics angle, active contour is defined as a parametric curve v (s)=[x (s), the y (s)] on the plane of delineation, s ∈ [0,1], and its energy function is written as:

$E_{\mathrm{snake}}={\∫}_0^1{E}_{\mathrm{snake}}\left(v\left(s\right)\right)\mathrm{ds}={\∫}_0^1(E_{\mathrm{int}}\left(v\left(s\right)\right)+E_{\mathrm{image}}\left(v\left(s\right)\right)+E_{\mathrm{con}}\left(v\left(s\right)\right))\mathrm{ds}---\left(1\right)$

Wherein: E _IntBe internal energy, be used for keeping continuity and the flatness of active contour at deformation process; E _ImageBe image energy, derive from the feature of image self; E _ConBe bound energy, produce by external force.E _ImageWith E _ConBe combined into external energy, be used to induce active contour to move, be designated as towards the border of area-of-interest:

E _ext＝E _image+E _con (2)

For obtaining an optimum matching on active contour and area-of-interest border, active contour need constantly change and adjust self shape in deformation process, to reach the minimized purpose of energy function, that is:

$v\′\left(s\right)=\arg \underset{s\∈\left[\mathrm{0,1}\right]}{\min }{E}_{\mathrm{snake}}\left(v\left(s\right)\right)---\left(3\right)$

Yet, in traditional deformation model not to bound energy E _ConClearly definition, only the external energy that is made of image energy is not enough to order about active contour and converges to the area-of-interest with depression border fully.Consider that area-of-interest has significant color characteristics and out-of-shape usually in the WEB image, the present invention redefines image energy E with colouring information _ImageExpression formula, and utilize the borderline unique point of area-of-interest structure bound energy E _Con, make active contour under the acting in conjunction of internal energy, image energy and bound energy in image the border of area-of-interest move.The discrete form of energy function is in the new deformation model:

$E_{\mathrm{snake}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}(\mathrm{\α}{E}_{\mathrm{int}1}\left(i\right)+\mathrm{\β}{E}_{\mathrm{int}2}\left(i\right)+\mathrm{\γ}{E}_{\mathrm{image}}\left(i\right)+\mathrm{\δ}{E}_{\mathrm{con}}\left(i\right))---\left(4\right)$

Wherein: N is the number at reference mark on the active contour; E _Int1Be the elastic energy in the internal energy, i.e. E _Int1=| v _i-v _I-1| ²/ 2; E _Int2Be the rigidity energy in the internal energy, i.e. E _Int2=| v _I-1-2v _i+ v _I+1| ²/ 2; E _ImageBe the image energy in the external energy, define by formula (5); E _ConBe the bound energy in the external energy, by formula (8) or formula (20) definition.α, beta, gamma and δ are weights, are used to adjust the contribution margin of each energy term to the integral energy function.

1) definition image energy

The WEB image all is coloured image basically, and the source is extremely variation also, and this just makes and can run into some inevitably because the pseudo-border situation that shade or high light produce when the WEB image is cut apart.Since colourity be a kind of shade, cover and the influence of factor such as Gao Guang under, still can be independent of the color characteristic of target shape and viewing angle, replace gray scale to calculate Grad with colourity among the present invention, can make the deformation model border of in cutting procedure, telling truth from falsehood.With color gradient definition image energy E _ImageFor:

$E_{\mathrm{image}}=-|\▿C_H(x,y)|---\left(5\right)$

Wherein: C _H(x, y) expression chromatic diagram picture; Be gradient operator.

In view of color space Lab is a kind of color space near human perception, and can measure color similarity easily with Euclidean distance, the present invention calculates colourity in the Lab of color space and texture (extracts colourity from a component and b component; From the L component, extract texture).The colourity image C _H(x y) utilizes the colourity gradient calculation colourity profile of image.Because colourity is defined on the annular region of [0,2 π], the calculating of colourity gradient should be measured with the angular distance between the different colourities.If h ₁And h ₂Be two chromatic values, the angular distance d (h between them ₁, h ₂) be defined as:

d(h ₁，h ₂)＝arccos(cosh ₁cosh ₂+sinh ₁sinh ₂) (6)

Utilize Canny operator detection boundaries point then, wherein formula (6) is adopted in the calculating of colour difference, obtains the gradient component on directions X respectively

With the gradient component on the Y direction

So (x, y) gradient magnitude of locating the reference mark is in the position $\sqrt{{C_{H_x}}^2(x,y)+{C_{H_y}}^2(x,y)}$ , gradient direction is

2) structure bound energy

The adding purpose of bound energy is extremely helpful to the structure bound energy in order to solve the depression border convergence problem of area-of-interest in the image, excavate shape facility and the movement tendency put on the border.Suppose v _I-1, v _i, v _I+1Be three successive control points on the active contour, with current reference mark v _iPoint to triangle Δ v _I-1v _iv _I+1Heart O _iVector define curvature, and the constraining force that produces by triangle heart gravitation of structure:

$F_{\mathrm{con}}\left(i\right)={\mathrm{\κ}}_i=\stackrel{\→}{v_i{o}_i}---\left(7\right)$

Use vector then

Length (mould) definition constraining force F _ConThe energy size that produces:

$E_{\mathrm{con}}\left(i\right)=\left|\stackrel{\→}{v_i{o}_i}\right|---\left(8\right)$

Can draw: reference mark v on the active contour _iAt position (x _i, y _i) to locate curvature big more, expression is by v _I-1, v _i, v _I+1The degree of crook of 3 curves that surround is big more, vector

Length just big more, v then _iThe bound energy that is subjected to coming from triangle heart gravitation is also big more; Otherwise curvature is more little, and the degree of crook of curve is more little, vector

Length just more little, the gravitation that is subjected to the triangle heart is also more little.

Utilize leg-of-mutton vector form to find the solution vector below

Value.Order and triangle Δ v _I+1v _iv _I+1Article three, three vector of unit length that edge direction is identical are respectively:

$\stackrel{\&RightArrow;}{e_1}=\stackrel{\&RightArrow;}{v_{i-1}{v}_i}/\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|,\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">e</figure-callout>}}_2}=\stackrel{\&RightArrow;}{v_i{v}_{i+1}}/\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|,\stackrel{\&RightArrow;}{e_3}=\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}/\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|---\left(9\right)$

Because leg-of-mutton heart is the intersection point of each interior angular bisector of triangle, the vector form definition of utilization angular bisector:

$\stackrel{\&RightArrow;}{v_{i-1}{o}_i}={\mathrm{\&lambda;}}_1(\stackrel{\&RightArrow;}{e_1}-\stackrel{\&RightArrow;}{e_3}),{\mathrm{\&lambda;}}_1\&GreaterEqual;0---\left(10\right)$

$\stackrel{\&RightArrow;}{v_i{o}_i}={\mathrm{\&lambda;}}_2(\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">e</figure-callout>}}_2}-\stackrel{\&RightArrow;}{e_1}),{\mathrm{\&lambda;}}_2\&GreaterEqual;0---\left(11\right)$

$\stackrel{\&RightArrow;}{v_{i+1}{o}_i}={\mathrm{\&lambda;}}_3(\stackrel{\&RightArrow;}{e_3}-\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">e</figure-callout>}}_2}),{\mathrm{\&lambda;}}_3\&GreaterEqual;0---\left(12\right)$

Wherein: λ ₁, λ ₂, λ ₃Be coefficient, then:

$\stackrel{\&RightArrow;}{v_{i-1}{v}_i}=\stackrel{\&RightArrow;}{v_{i-1}{o}_i}-\stackrel{\&RightArrow;}{v_i{o}_i}\&DoubleLeftRightArrow;\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|\&CenterDot;\stackrel{\&RightArrow;}{e_1}={\mathrm{\&lambda;}}_1(\stackrel{\&RightArrow;}{e_1}-\stackrel{\&RightArrow;}{e_3})-{\mathrm{\&lambda;}}_2(\stackrel{\&RightArrow;}{e_2}-\stackrel{\&RightArrow;}{e_1})---\left(13\right)$

$\stackrel{\&RightArrow;}{v_i{v}_{i+1}}=\stackrel{\&RightArrow;}{v_i{o}_i}-\stackrel{\&RightArrow;}{v_{i+1}{o}_i}\&DoubleLeftRightArrow;\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|\&CenterDot;\stackrel{\&RightArrow;}{e_2}={\mathrm{\&lambda;}}_2(\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">e</figure-callout>}}_2}-\stackrel{\&RightArrow;}{e_1})-{\mathrm{\&lambda;}}_3(\stackrel{\&RightArrow;}{e_3}-\stackrel{\&RightArrow;}{e_2})---\left(14\right)$

$\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}=\stackrel{\&RightArrow;}{v_{i+1}{o}_i}-\stackrel{\&RightArrow;}{v_{i-1}{o}_i}\&DoubleLeftRightArrow;\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|\&CenterDot;\stackrel{\&RightArrow;}{e_3}={\mathrm{\&lambda;}}_3(\stackrel{\&RightArrow;}{e_3}-\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">e</figure-callout>}}_2})-{\mathrm{\&lambda;}}_1(\stackrel{\&RightArrow;}{e_1}-\stackrel{\&RightArrow;}{e_3})---\left(15\right)$

Again because:

$\stackrel{\&RightArrow;}{v_{i-1}{v}_i}+\stackrel{\&RightArrow;}{v_i{v}_{i+1}}+\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}=0\&DoubleLeftRightArrow;\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|\&CenterDot;\stackrel{\&RightArrow;}{e_1}+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|\&CenterDot;\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">e</figure-callout>}}_2}+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|\&CenterDot;\stackrel{\&RightArrow;}{e_3}---\left(16\right)$

By (13), (14), (15), (16) can get:

$\left\{\begin{array}{c}{\mathrm{\&lambda;}}_1=(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|\&CenterDot;\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|)/(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|)\\ {\mathrm{\&lambda;}}_2=(\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|\&CenterDot;\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|)/(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|)\\ {\mathrm{\&lambda;}}_3=(\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|\&CenterDot;\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|)/(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|)\end{array}\right.---\left(17\right)$

So formula (11) can be rewritten as:

$\stackrel{\&RightArrow;}{v_i{o}_i}={\mathrm{\&lambda;}}_2(\stackrel{\&RightArrow;}{{\mathrm{<figure-callout\; id="2"\; label="e"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">e</figure-callout>}}_2}-\stackrel{\&RightArrow;}{e_1})=((\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|\&CenterDot;\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|)/(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|))(\stackrel{\&RightArrow;}{v_i{v}_{i+1}}/\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|-\stackrel{\&RightArrow;}{v_{i-1}{v}_i}/\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|)---\left(18\right)$

$=(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|\&CenterDot;\stackrel{\&RightArrow;}{v_i{v}_{i+1}})/(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|)-(\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|\&CenterDot;\stackrel{\&RightArrow;}{v_{i-1}{v}_i})/(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|)$

In formula (18), $\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|/(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|)$ With $\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|/(\left|\stackrel{\&RightArrow;}{v_{i-1}{v}_i}\right|+\left|\stackrel{\&RightArrow;}{v_i{v}_{i+1}}\right|+\left|\stackrel{\&RightArrow;}{v_{i+1}{v}_{i-1}}\right|)$ Be constant, might as well be made as k ₁With k ₂, the substitution formula in (18) is:

$\stackrel{\&RightArrow;}{v_i{o}_i}=k_1\stackrel{\&RightArrow;}{v_i{v}_{i+1}}-k_2\stackrel{\&RightArrow;}{v_{i-1}{v}_i}$

$=k_1((x_{i+1}-x_i),(y_{i+1}-y_i))-k_2((x_i-x_{i-1}),(y_i-y_{i-1}))$

$=(k_1(x_{i+1}-x_i),k_1(y_{i+1}-y_i))-(k_2(x_i-x_{i-1}),k_2(y_i-y_{i-1}))---\left(19\right)$

$=(k_1(x_{i+1}-x_i)-k_2(x_i-x_{i-1}),k_1(y_{i+1}-y_i)-k_2(y_i-y_{i-1}))$

$=(k_1{x}_{i+1}-(k_1+k_2)x_i+k_2{x}_{i-1},k_1{y}_{i+1}-(k_1+k_2)y_i+k_2{y}_{i-1})$

Draw reference mark v by above-mentioned derivation _iAt vector

The size of the bound energy on the direction is $\sqrt{{(k_1{x}_{i+1}-(k_1+k_2)x_i+k_2{x}_{i-1})}^2+{(k_1{y}_{i+1}-(k_1+k_2)y_i+k_2{x}_{i-1})}^2}.$

Consider when three continuous on active contour reference mark conllinear can't constitute triangle, the bound energy that is produced by triangle heart gravitation equals zero, the present invention increases three point on a straight line and judges, augments external force (from the acting force between the adjacent reference mark) if three point on a straight line is judged to set up then to add.Suppose that active contour center c is at vector The acting force that produces on the direction is F _Ci, utilize the angle of cosine law compute vectors, then directed force F _CiAt vector

The energy that produces on the direction is:

$E_{\mathrm{con}}\left(i\right)=f_i\cos (\stackrel{\&RightArrow;}{{\mathrm{cv}}_i},\stackrel{\&RightArrow;}{{\mathrm{cv}}_{i+1}})=(f_i/\left(\left|\stackrel{\&RightArrow;}{{\mathrm{cv}}_i}\right|\left|\stackrel{\&RightArrow;}{{\mathrm{cv}}_{i+1}}\right|\right))(\stackrel{\&RightArrow;}{{\mathrm{cv}}_i}\&CenterDot;\stackrel{\&RightArrow;}{{\mathrm{cv}}_{i+1}})---\left(20\right)$

Wherein: f _iBe directed force F _CiValue; Be inner product.In formula (20),

Be constant, might as well be made as k ₃, the substitution formula in (20) is:

$E_{\mathrm{con}}\left(i\right)=k_3(\stackrel{\&RightArrow;}{{\mathrm{cv}}_i}\&CenterDot;\stackrel{\&RightArrow;}{{\mathrm{cv}}_{i+1}})=k_3((x_i-x_c)(x_{i+1}-x_c)+(y_i-y_c)(y_{i+1}-y_c))---\left(21\right)$

Wherein: the position of active contour center c is (x _c, y _c), order $x_c=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{x}_i/N,y_c=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{y}_i/N.$

2, the coarse segmentation of area-of-interest:

In the Lab of color space, extract color and texture information, utilize automatic cluster to serve as initialization operation, realize coarse segmentation area-of-interest in the image, specific as follows:

For the image that comes from the WEB environment, it is difficult directly with deformation model the area-of-interest in the image accurately being cut apart, and needs to add one and can determine that regional approximate location can conveniently obtain the initialization operation of active contour initial position in the model again.Because this operation do not need to obtain accurate segmentation effect, mainly be that area-of-interest is split from background, so more emphasis be lower calculation cost.The present invention adopts the coarse segmentation that realizes area-of-interest based on the clustering method of image subblock.At first image is divided into size and is the disjoint sub-piece of M * M, then each sub-piece is extracted characteristics of image, carry out automatic cluster with the k-means algorithm again.The integrality of choosing to guarantee area-of-interest of sub-block size is a prerequisite, and when the group block size was big more, computing time was short more, and the feature of extracting will be coarse more; Otherwise, the group block size more hour, the feature of extraction is careful more, but calculated amount will increase.For compromise calculations speed and characteristic validity, use 4 * 4 sub-block size in the present invention as default value.

1) extracts characteristics of image

At each image subblock i, color is the averages of three color components in image subblock for three-dimensional in the characteristics of image that extracts in the Lab of color space, is designated as:

With

Textural characteristics also is three-dimensional, is through the square root of wavelet transformation (wavelet transformation is very effective to the Analysis of Complex texture) back three high frequency band HL, LH, the last energy of HH, supposes that the coefficient on the HL frequency band is: { hl _{I, j}, hl _{I, j+1}), hl _{I+1, j}, hl _{I+1, j+1}; Coefficient on the LH frequency band is: { lh _{I, j}, lh _{I, j+1}, lh _{I+1, j}, lh _{I+1, j+1}; Coefficient on the HH frequency band is:

{ hh _{I, j}, hh _{I, j+1}, hh _{I+1, j}, hh _{I+1, j+1}, then three textural characteristics are expressed as respectively:

$f_i^4={(({{\mathrm{hl}}_{i,\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">j</figure-callout>}}}^2+{{\mathrm{hl}}_{i,j+1}}^2+{{\mathrm{hl}}_{i+1,\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">j</figure-callout>}}}^2+{{\mathrm{hl}}_{i+1,j+1}}^2)/4)}^{1/2};$ $f_i^5={(({{\mathrm{lh}}_{i,\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">j</figure-callout>}}}^2+{{\mathrm{lh}}_{i,j+1}}^2+{{\mathrm{lh}}_{i+1,\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">j</figure-callout>}}}^2+{{\mathrm{lh}}_{i+1,j+1}}^2)/4)}^{1/2};$ $f_i^6={(({{\mathrm{hh}}_{i,\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">j</figure-callout>}}}^2+{{\mathrm{hh}}_{i,j+1}}^2+{{\mathrm{hh}}_{i+1,\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="H2008101627467E00021.png"\; state="\{\{state\}\}">j</figure-callout>}}}^2+{{\mathrm{hh}}_{i+1,j+1}}^2)/4)}^{1/2}.$

2) carry out automatic cluster

The k-means algorithm is a kind of quick clustering method that is proposed by MacQueen, utilizes this algorithm that image subblock is gathered in each bunch according to the sextuple feature of extracting among the present invention, reaches the purpose that image segmentation is become zones of different with this.Because the k-means algorithm is a kind of nothing supervision clustering algorithm, bunch number need specify in conjunction with priori, adopt a kind of method of adaptively selected bunch of number in the present invention, promptly to guarantee that the target integrality is a prerequisite, the number that increases step by step bunch from value 2 beginning is till prerequisite can not satisfy.

3, obtaining automatically of active contour initial position:

Cluster result is carried out binary conversion treatment, produce the two-value mask figure that only constitutes, adopt the generation rays method on the foreground area border, to choose the reference mark, obtain the initial position of active contour in the deformation model by prospect and background, specific as follows:

1) generates two-value mask figure

Because the difference of the complexity of image to be split, cause that automatic cluster obtains bunch number also different.Among the present invention owing to only need area-of-interest and background area are separated, so for simplicity, bunch number what are all need pass through binary conversion treatment, i.e. the two-value mask figure that acquisition only is made of prospect and background.Operating process is: the mask figure that generates an identical size earlier for original image, then each pixel in the original image is handled, if belong to the point of (comprising on the area-of-interest border) in the area-of-interest, then in mask figure on the correspondence position value of setting be 1 (prospect), if belong to the overseas point of region of interest, all regard background as, then in mask figure on the correspondence position value of setting be 0 (background), the pixels all up to original image all dispose.

2) obtain the initial position of active contour

Utilize the generation rays method on the foreground area border, to choose the reference mark, obtain the initial position of active contour:

A) bunch center with area-of-interest among the two-value mask figure is a starting point, structure L bar ray l _i(i=1,2 ..., L), the angle between two adjacent rays is 2 π/L.

B),, be the frontier point of area-of-interest from bunch pixel that center searching value changes for every ray.

C) successively L bar ray is carried out the operation of step c), with L the frontier point that the obtains closed curve of formation that is linked in sequence.

D) curve that obtains is above carried out spline interpolation, obtain the number (generally setting two maximum spacing and minimum spacings between the reference mark) at reference mark on the suitable active contour, this curve is the initial position of the active contour that carries out deformation.

4, two of deformation model sections execution:

Adopt two sections implementation strategies to implement the concrete application of deformation model, obtain the accurate segmentation result of area-of-interest in the image, specific as follows:

Two sections implementation strategies are the methods that execution speed is cut apart in a kind of quickening of adopting when realizing deformation model.

Because deformation model is all wanted the curvature value at each reference mark on the computational activity profile in each iteration, need be directly proportional with the reference mark number computing time of cost.Secondly, add bound energy prematurely in deformation model, if do not adjust the value of parameter, the speed of convergence of active contour can be slack-off in the time of a little less than external force is crossed; Otherwise when external force is crossed when strong, even other zone except that depression has all converged to correct border, along with the active contour continuation approaches to sunk area, the segment of curve that correctly arrives the area-of-interest boundary position also can be crossed the border, the convergent phenomenon occurs.Consider above these situations, propose two sections implementation strategies.

1) first section: utilize traditional deformation model can better converge to non-sunk area and do not need to increase the advantage of the extra cost of calculation control point curvature, make active contour converge to a scope preferably rapidly earlier.

2) second section: add bound energy from triangle heart gravitation (active contour center gravitation is done and augmented).Sunk area falls into deeply more, and bound energy just adds greatly more, drives the border that active contour rapidly converges to sunk area.

2 explanations:

A) division of two sections iterationses is mainly decided according to the border characteristics of area-of-interest, for example: overall more smooth when the zone boundary, when one to two depression or protrusion are arranged, can suitably amplify first section ratio; When the zone boundary relative complex, the depression of appearance or protrusion can suitably amplify second section ratio more for a long time.If division proportion will be set automatically, compromise method is to divide half-and-half.

What b) sampling at reference mark was adopted on the active contour is double sampling mechanism, does not promptly stipulate the reference mark number of taking a sample in each iteration, and just stipulates the minimum and maximum spacing between the reference mark, to increase adaptively and to delete the reference mark.Consider that spacing and splitting speed, processing accuracy are relevant, spacing is big more, the reference mark number more less, splitting speed is fast more, it is just coarse more to handle; On the contrary, spacing is more little, the reference mark number is many more, splitting speed is slow more, it is just meticulous more to handle, so adopt the bigger quick convergence of spacing in first section; Adopt the less slow convergence of spacing in second section.The benefit of doing like this is when active contour during near the border of area-of-interest, allows active contour to do slowly and moves, and obtains more meticulous segmentation effect.The method that two sections strategies of this usefulness are realized, make that segmentation effect is better, speed is also faster, the occasion that not only is fit to weak boundary, and when running into the zone boundary that to cut apart when very complicated, also can obtain gratifying segmentation result, make deformation model have more dirigibility and practicality by adjusting every section iterations ratio, change reference mark sampling spacing and expanding hop count.

For verifying the validity of the dividing method that proposes among the present invention, design the segmentation result that four groups of examples of implementation are used for comparison deformation model of the present invention and other model.Used experimental subjects has composograph, also the true picture that comes from Corel image library and the WEB environment is arranged.Experimental situation is a 1.86GHZ T2350 CPU 2G internal memory, and all programs realize with the MATLAB7.0 programming.

Embodiment 1: Fig. 2 is the segmentation result comparisons of various deformation models to the sunk area object.The image size

Be 64 * 64.((a) traditional deformation model (α=1, β=0.2, γ=10, δ=0) (200 iteration); (b) balloon power model (α=1, β=0.2, γ=10, δ=0.05) (200 iteration); (c) apart from Potential Model (α=0.05, β=0, γ=0.5, δ=0) (200 iteration); (d) GVF model (α=0.05, β=0, γ=1, δ=0.5) (200 iteration); (e) improved model among the present invention (α=1, β=0.2, γ=10, δ=4) (180 iteration); (f) two of the improved model among the present invention sections implementation strategies (α=1, β=0.2, γ=10, δ=0; δ=4) (140 iteration)).

The unification between the different models had both been considered in the selection of the weights of each energy term in the deformation model, guaranteed as far as possible that also every kind of model can both obtain boundary segmentation effect preferably in the iterations of regulation, so the constantly debugging of determining in force of final weights obtains.The initial position of active contour obtains by drawing circule method.As can be seen from Fig. 2, (a) all can not correctly cut apart the border of sunk area object to (c).Although (d) obtain to plant the better boundary segmentation result of model,, can not in less iterations, converge to correct border for dark sunk area active contour owing to have the limited defective of convergence range equally than first three.(e) and (f) can both correctly cut apart the border of sunk area object, (e) need be through 180 iteration, and (f) only need just obtain correct segmentation result through 140 iteration, execution speed has been accelerated greatly.

Embodiment 2: Fig. 3 is the segmentation result comparisons (275 iteration) of various deformation models to area-of-interest in the image.Image derives from the Corel image library, and the image size is 100 * 100.((a) former figure; (b) traditional deformation model (α=1, β=0, γ=6, δ=0); (c) balloon power model (α=0.6, β=0, γ=2, δ=0.15); (d) apart from Potential Model (α=0.05, β=0, γ=0.5, δ=0); (e) be only with the model (α=1, β=0, γ=8, δ=3) that produces from the bound energy of triangle heart gravitation; (f) for adopting deformation model (α=1, β=0, γ=6, δ=0 of two sections implementation strategies among the present invention; δ=3)).

The initial position of active contour obtains by drawing circule method in the present embodiment.Utilize the colourity gradient to form the colourity profile of area-of-interest (" the flower zones of image central authorities ") in the image.As can be seen from Fig. 3, because the border less complex of area-of-interest, (b) can both cut apart the approximate bounds position that obtains area-of-interest in the image to (e), but it is accurate not enough to cede territory in the protrusion punishment on the recess of the similar cusp on the regional left side and the right, does not reach higher segmentation precision.(f) according to the border characteristics of area-of-interest to be split in the image, first section iterations ratio with second section is divided into 2/3 and 1/3 in; The reference mark spacing is set in first section simultaneously, and to be 2 minimums to the maximum be to be provided with in 0.8, the second section to get the reference mark spacing to be 1 minimum to the maximum be 0.2, and the segmentation result that obtains in this case is best.

Embodiment 3, and: Fig. 4 is the deformation model that proposes among the present invention when getting different image subblocks relatively (iterations: first section is 100 to the segmentation result of area-of-interest in the image; Second section is 50).(former figure is the same with embodiment 2; (a) cluster of sub-block size=2 * 2; (b) cluster of sub-block size=4 * 4; (c) cluster of sub-block size=5 * 5; (d) be the segmentation result corresponding with (a); (e) be and (b)) corresponding segmentation result; (f) be the segmentation result corresponding with (c); (α=1, β=0, γ=6, δ=0; δ=3)).

The initial position of active contour is obtained by automatic cluster in the present embodiment.As can be seen from Fig. 4, although cluster result is different under three kinds of situations (acquisition of (a) than (b) and (c) all good cluster result), but behind the deformation model that adopts two sections strategies, in 150 times identical iterationses, obtained the essentially identical segmentation result of levels of precision ((d) to (f)).This shows, the area-of-interest in the image can be split basically that the quality quality of cluster result is irrelevant with the execution performance of deformation model as long as the method among the present invention can guarantee cluster operation.Yet, owing to the execution time of automatic cluster can influence whole splitting speed.For example: in the Lab of color space to present embodiment in used image size be that 100 * 100 coloured image is realized the automatic cluster based on color and textural characteristics, (a) execution time is 3.547 seconds; (b) execution time is 1.078 seconds; (c) execution time is 0.593 second.Therefore, when needs are handled large batch of WEB image, under the prerequisite that does not influence the region of interest domain integrity, in cluster, should get bigger slightly image subblock as far as possible, to obtain cutting procedure fast.

Embodiment 4: Fig. 5 is the influence of the acquisition methods of different active contour initial position to various deformation models.Image is downloaded from the WEB environment, and size is 100 * 100.((a) former figure; (b) traditional deformation model (α=1, β=0.2, γ=8, δ=0); (c) balloon power model (α=1, β=0.2, γ=2, δ=0.08); (d) for adopting deformation model (α=1, β=0.2, γ=8, δ=0 of two sections implementation strategies among the present invention; δ=3.6); (e) for adopting deformation model (α=1, β=0.2, γ=8, δ=0 of two sections implementation strategies among the present invention; δ=3), wherein (b) comes from the rectangular area (350 iteration) of automatic generation to the initial position of the active contour of (d); The initial position of active contour (e) comes from automatic cluster (150 iteration)).

As can be seen from Fig. 5, not only active contour does not converge to the borders of area-of-interest (" rabbits of image central authorities ") (b), also has been subjected to the attraction of certain Local Extremum in the image upper left corner.(c) active contour is not subjected to the interference of Local Extremum in, has obtained zone boundary substantially accurately, but at some positions, and the gap location between the recess that is connected with health such as the head of rabbit, two pin does not approach real border fully.(d) segmentation result for obtaining with two sections strategy execution deformation models among the present invention.Consider among this embodiment the area-of-interest border than the complexity among the embodiment 2, expand original two sections, the deformation process in original second section is decomposed once more, in 350 iterative process, first section accounts for 3/7 (150 times), and the maximum spacing of reference mark sampling is 4, minimum spacing is 2; Second section (I) accounts for 3/7 (150 times), and the maximum spacing of reference mark sampling is 2, minimum spacing is 0.5; Second section (I) accounts for 1/7 (50 times), and the maximum spacing of reference mark sampling is 0.5, minimum spacing is 0.2.Active contour not only approaches the gap between the head of rabbit and health junction, two pin exactly, and even the outshot of rabbit face is also split, has really realized accurately cutting apart of area-of-interest.(e) provided elder generation in and obtained the initial position of active contour, again with two sections results that strategy enforcement is cut apart of deformation model by automatic cluster.Owing to utilize the color of image and the cluster operation of textural characteristics realization to obtain the active contour initial position of a basic access areas real border, so only just obtained the same accurate segmentation result with (d) with 150 iteration.This shows that the deformation model that proposes among the present invention can effectively improve segmentation precision, the iterations of the profile that takes in sail; If use automatic cluster, can make deformation model not be subjected to the influence of active contour initial position and the interference of Local Extremum simultaneously, also can shorten the whole execution time that area-of-interest is cut apart in the image as initialization.

Claims (3)

1. the dividing method of area-of-interest in the WEB image is characterized in that comprising following content:

1) utilizing color gradient definition image energy, is auxilliary structure bound energy based on triangle heart gravitation, active contour center gravitation, sets up the deformation model that meets the WEB picture characteristics;

2) in the Lab of color space, extract color and texture information, utilize automatic cluster to serve as initialization operation, realize coarse segmentation area-of-interest in the image;

3) cluster result is carried out binary conversion treatment, produce the two-value mask figure that only constitutes, adopt the generation rays method on the foreground area border, to choose the reference mark, obtain the initial position of active contour in the deformation model by prospect and background;

4) adopt two sections implementation strategies to implement the concrete application of deformation model, obtain the accurate segmentation result of area-of-interest in the image;

Described foundation meets the deformation model of WEB picture characteristics, setting up in the process of deformation model, and the structure energy function $E_{\mathrm{snake}}={\&Integral;}_0^1{E}_{\mathrm{snake}}\left(v\left(s\right)\right)\mathrm{ds}={\&Integral;}_0^1(E_{\mathrm{int}}\left(v\left(s\right)\right)+E_{\mathrm{image}}\left(v\left(s\right)\right)+E_{\mathrm{con}}\left(v\left(s\right)\right))\mathrm{ds};$ Wherein: active contour is defined as a parametric curve v (s)=[x (s), the y (s)] on the plane of delineation, s ∈ [0,1]; E _IntBe internal energy, be used for keeping continuity and the flatness of active contour at deformation process; E _ImageBe image energy, derive from the feature of image self; E _ConBe bound energy, produce, utilize chrominance information definition image energy E by external force _Image=-| ▽ C _H(x, y) |; Wherein: C _H(x, y) expression chromatic diagram picture; ▽ is a gradient operator, utilizes reference mark v _iPoint to triangle Δ v _I-1v _iv _I+1Heart o _iVector definition curvature, obtain the bound energy that produces by triangle heart gravitation

And increase the three point on a straight line judgement, obtain coming from the auxiliary bound energy of active contour center gravitation

Suppose that active contour center c is at vector

The acting force that produces on the direction is F _Ci, f _iBe directed force F _CiValue; Be inner product;

Two sections implementation strategies of described employing are implemented the concrete application of deformation model, in the implementation process of two sections implementation strategies of deformation model, do not add bound energy in first section, to reduce the extra cost of calculating curvature, utilize image energy to order about the active contour convergence; Add bound energy in second section, obtaining the global convergence of active contour, and in first section with the bigger quick convergence of reference mark sampling spacing; With the less slow convergence of sampling spacing, area-of-interest accurately cuts apart in the realization image in second section.

2. the dividing method of area-of-interest in a kind of WEB image according to claim 1, it is characterized in that: described realization is to the coarse segmentation of area-of-interest in the image, in the coarse segmentation process of area-of-interest, utilize color space Lab in i image subblock, to extract color characteristic f _i ¹, f _i ²And f _i ³Utilize wavelet transformation texture feature extraction f _i ⁴, f _i ⁵And f _i ⁶, and utilize the k-means algorithm image subblock of M * M to be distinguished in different bunches according to the sextuple feature of prior extraction, bunch number select to adopt adaptive mode.

3. the dividing method of area-of-interest in a kind of WEB image according to claim 1, it is characterized in that: the described initial position that obtains active contour in the deformation model, in the automatic acquisition process of active contour initial position, cluster result is carried out binary conversion treatment, produce the two-value mask figure that only constitutes by prospect and background, and be starting point with the center of prospect region, outwards generate L bar ray, angle between adjacent two rays is 2 π/L, and the curve that is formed by connecting by the intersection point on these rays and foreground area border constitutes the initial position of active contour in the deformation model.

Images