CN101673345B - Method for extracting target closed contour based on shape prior - Google Patents

Abstract

The invention discloses a method for extracting target closed contour based on shape prior, belonging to the field of computer application technology. From the point of view of simitating the information of human vision, the extraction method inhibiting salient edges of noise and texture is adopted, the shape prior is blended on the basis, and a new target closed contour extraction algorithm can be provided by utilizing curve evolution technology based on differential geometry, so that the important technology is provided for automatic target detection. The experiments on a plurality of image libraries prove that when the image background is very disordered and has partial shading, and part of the target boundary is indistinct caused by weaker light, the method can obtain the whole target closed contour. Simultaneously, when only a few of templates of a certain category of objects are provided, the method can extract the contour of the category of objects with different postures, namely, the method has invariance property for elastic deformation to a certain extent, so as to be directly applied to intelligent image processing. The method can accurately extract the target closed contour and has wide applicability.

Description

Method based on the extraction target closed contour of shape prior

Technical field

The invention belongs to the Computer Applied Technology field, particularly a kind of method of the extraction target closed contour based on shape prior.

Background technology

People are in the face of a strange and complex image time, and people can mark objective contour rapidly, exactly.Yet, when quite simple thing is given computing machine and done concerning people this, be not so easy just.Having lacked information such as priori, where computing machine is the edge, where be objective contour if can't being judged, which is the marking area that we want, and which content is redundant noise.Therefore, the objective contour in the complicated irregular image extracts and has just become the problem of often facing in the computer vision.

Extracting target closed contour is one of approach of target detection and tracking; Although the method for some non-supervision has obtained local success on image-forming condition better image (for example: background is single, the target background contrast strongly, is not blocked etc.), from the complex natural scene, obtain the target integrity profile that possibly have arbitrary shape and be still a still unsolved difficult problem.

What accomplish that the difficulty of this task forms sharp contrast with computing machine is that the human visual system has high efficiency and robustness when the perception objective contour.How could let computing machine can find objective contour rapidly accurately as human? The present invention concentrates the target closed contour extraction algorithm under the inspiration of research biological vision, for automatic target detects important technology is provided.Existing contour extraction method is divided into two types: based on the method and the driving wheel contour method of cluster.

Method based on cluster comprises the k-mean cluster, conviction transmission (Affinity propagation), and the most general spectral clustering of in profile extracts, using." spectrum " is meant the energy spectrum of the data that the second-order statistic of data dependence is represented.Spectral clustering at first makes up relational matrix according to the relation between the picture element, and the proper vector of the analysis matrix data that will constitute matrix are divided into different classes then, makes the just in time profile of a target of correspondence of each type.Minimum cutting (mini-cut) [Wu 1993], average cutting (averagecut) [Sarkar 2000] and normalization cutting (normalized cut) [Shi 1997] are cut apart and can be divided into to number according to the Different function forms of relational matrix and the proper vector of getting with order (possibly get maximum proper vector and also possibly get the minimal characteristic vector) figure.

Shortcoming based on the contour extraction method of cluster is: most applications can produce discontinuous profile fragment, but not a complete closed contour; This unsupervised method is relatively poor to the robustness of noise; The computation complexity of spectral clustering method is higher.

The driving wheel contour method is through under the acting in conjunction of image force and external constraint power, and the motion of controllable continuous deformation curve obtains the final goal profile.People such as Kass have proposed active contour model (Active ContourModel, snake model Snake Model is otherwise known as) in 1987.The energy equation of model can be made up of two parts: a part defined one telescopic and have an internal energy of the curve of certain hard degree; Controlling the physical behavio(u)r and the local continuity of model, making curve keep level and smooth again when curve is inwardly shunk.Another part is an external energy, is made up of image energy quantifier and external force constraint energy term again.By defining suitable image energy item, can make curve to required characteristics of image place motion.The pairing external constraint power of external constraint energy can artificially increase, and is used for strengthening the constraint to curve.

Active contour model has the incomparable advantage of previous methods:

Snake model provides a kind of effectively unified settling mode for the multiple visual problem that needed to solve respectively in the past.Edge, line segment and objective contour have same significance under this model mechanism; Motion tracking and matching problem can easily solve under Unified frame too.

Snake model has changed the single representation mode of lower-level vision information in the past; Through the extreme value of one group of separate lower-level vision problem is provided; With diversified mode to the high-level vision transmission; Mode and high-level vision knowledge through iteration are carried out alternately, and both are unified in the same characteristic extraction procedure first.

After suitable initialization, model can converge on the minimal value place of energy equation automatically.

Meanwhile, active contour model also has intrinsic defective:

Responsive to initial position, need to rely on manual or other machine-processedly are placed on initial curve near the interested characteristics of image, so that obtain correct convergence result.

Because the non-convexity of energy function causes the snake model might be retracted to Local Extremum, even disperses.

Exactly because the plurality of advantages that active contour model had makes it in nearly more than 20 years, obtain swift and violent development.Be broadly divided into based on the active contour model at edge with based on two big types of the active contour models in zone, and introduced curve evolvement technology, come optimization aim with the mode of finding the solution PDE based on level set.

Its main thought is the zero level line that contour curve C is expressed as imbedding function (embedding function) φ, promptly C={ (x, y) | (t)=0}, the evolution of contour curve just is converted into the evolution of its corresponding imbedding function to φ like this, promptly for x, y $\frac{\∂C}{\∂t}\→\frac{\∂\mathrm{\φ}}{\∂t}.$ That first is applied to image segmentation with the level set technology is Caselles et al. (1993); He does a convolution with image and the Gaussian smoothing filter that with σ is variance earlier; Through defining suitable image energy quantifier and curvature bound term; Can make the edge motion of curve in image, and reduce the influence of noise as much as possible, can under complex background, extract complete closed contour difform target to the result.

This model can obtain satisfied profile when the contrast of the prospect of image and background is very strong, then powerless to discontinuous edge.Therefore not being suitable for the illumination dimness causes the not so good images of image-forming condition such as edge fog, partial occlusion.In addition, the input of directly extracting as profile with the image of making of the Gaussian wave filter after the smoothing processing, limited to the Noise Suppression effect, only the image to background comparison " totally " has effect preferably.Under complex background, might cause curve not restrained.Also restriction to some extent of shape to target if target contains the profile of depression, is difficult to be deep into the depression position during curve evolvement.

Summary of the invention

The problem that the present invention will solve is to disturb for be subject to ground unrest and the inner vein that overcomes prior art; Do not possess to block, the robustness of illumination and to the above-mentioned defective of elastically-deformable unchangeability; A kind of method of the extraction target closed contour based on shape prior is provided; It is characterized in that, comprise following steps:

Step 1 is utilized the imageing sensor input picture;

Step 2 to the image of step 1 input, is set up the bank of filters p of Simulation of Complex cell ₁(f ₁, θ ₁)～p _n(f _n, θ _n), and image carried out Filtering Processing, obtain the edge;

Step 3 is set up the texture inhibition function r that simulates non-classical receptive field ₁(f ₁, θ ₁)～r _n(f _n, θ _n);

Step 4 suppresses both synergy c to the edge extracting of step 2 Simulation of Complex cell and the texture of step 3 simulation non-classical receptive field ₁(f ₁, θ ₁)～c _n(f _n, θ _n) maximize processing, obtain prominent edge figure;

Step 5, the step of the similarity of shape among shape among calculation template 1～template m and the prominent edge figure;

Step 6, the step of the accurate closed contour of extraction target.

Described step 2,3,4 is the simulating human vision system, and performing step is following:

(1a) input picture is carried out filtering by the Gabor bank of filters of one group of odd symmetry and even symmetry; This is similar to the operation that the complex cell in the human primary visual cortex is taked input picture, purpose be from input picture, extract have CF and towards the edge line segment:

$p_{f,\mathrm{\θ}}=\sqrt{{\left(p_{f,\mathrm{\θ}}^e\right)}^2+{\left(p_{f,\mathrm{\θ}}^o\right)}^2}---\left(1\right)$

$p_{f,\mathrm{\θ}}^e=I(x,y)\⊗h^e(x,y,\mathrm{\θ},f)---\left(2\right)$

$p_{f,\mathrm{\θ}}^o=I(x,y)\⊗h^o(x,y,\mathrm{\θ},f)---\left(3\right)$

$h^e(x,y,\mathrm{\θ},f)=\exp (-\frac{1}{2}(\frac{{x^{\′}}^2}{{{\mathrm{\σ}}_x}^2}+\frac{{y^{\′}}^2}{{{\mathrm{\σ}}_y}^2}))\cos \left(2\mathrm{\πf}{x}^{\′}\right)---\left(4\right)$

$h^o(x,y,\mathrm{\θ},f)=\exp (-\frac{1}{2}(\frac{{x^{\′}}^2}{{{\mathrm{\σ}}_x}^2}+\frac{{y^{\′}}^2}{{{\mathrm{\σ}}_y}^2}))\sin \left(2\mathrm{\πf}{x}^{\′}\right)---\left(5\right)$

Wherein x '=xcos θ+ysin θ y '=-xsin θ+ycos θ,

I (x, y) expression one width of cloth be of a size of M * N image in (x, the monochrome information of y) locating; p ^eThe Gabor wave filter h of expression even symmetry ^eTo the result of image filtering, p ^oRepresent odd symmetric Gabor wave filter h ^oTo the result of image filtering, p _{F, θ}The energy summation of representing these two kinds of filter filtering results, the response of corresponding complex cell;

The expression convolution algorithm, exp () expression is the exponent arithmetic at the end with e; F and θ represent respectively wave filter frequency and towards; σ _x ²And σ _y ²Represent the width of wave filter on x direction and y direction respectively.

(1b) consider to connect the inhibiting effect that produces through side between the cell, calculating the energy that suppresses to produce is r _{F, θ}, purpose is to suppress noise edge and mixed and disorderly texture edge, thereby only keeps the edge of remarkable structure, this be the priority processing that in the long-term evolution process, forms of human visual system significantly or information of interest, ignore the strategy of redundant information:

$r_{f,\mathrm{\θ}}=p_{f,\mathrm{\θ}}\⊗\widetilde{\mathrm{\ω}}---\left(6\right)$

$\mathrm{\ω}(x,y)=\frac{1}{2\mathrm{\π}{\left(4\mathrm{\σ}\right)}^2}\exp (-\frac{x^2+y^2}{2{\left(4\mathrm{\σ}\right)}^2})-\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})---\left(7\right)$

$\widetilde{\mathrm{\ω}}=\frac{\mathrm{\ω}}{\left|\right|\mathrm{\ω}\left|\right|}$

In the formula, σ representes the bandwidth of wave filter, || || expression norm, r _{F, θ}The expression inhibiting effect is with the difference of gaussian wave filter of normalization Response p with complex cell _{F, θ}Convolution represent, be the non-classical receptive field pattern of simulation primary visual cortex cell.

(1c) from the edge that complex cell extracts, get rid of repressed edge, promptly obtain the edge that keeps

c _f，θ＝[p _f，θ-αr _f，θ] ₊ (8)

α is a constant, is used for regulating inhibiting power.[] ₊For getting the operation of nonnegative number, nonnegative number remains unchanged after operating through this, and negative is output as zero after operating through this.

(1d) when θ get different towards the time, from each c of correspondence _{F, θ}In get maximum as prominent edge.

sc _f＝max{c _f，θ|θ＝θ ₁，θ ₂，…，θ _n}(9)

Described step (5) realizes through the distance between the definition shape.Suppose s _iAnd s _jBe any two shapes, SD (s _i, s _j) represent the distance between them, weigh shape and the similarity of the contour shape among the prominent edge figure in each template with this;

$\mathrm{SD}(s_i,s_j)=\sqrt{{\left|\right|\mathrm{\φ}\left(s_i\right)-\mathrm{\φ}\left(s_j\right)\left|\right|}^2+{\left|\right|\▿\mathrm{\φ}\left(s_i\right)-\▿\mathrm{\φ}\left(s_j\right)\left|\right|}^2}---\left(10\right)$

Wherein, φ (s _i) be shape s _iThe symbolic distance function, (x y) just in time is positioned at s to any point in the image two dimensional surface _iProfile on the time, symbolic distance is 0; When (x y) is positioned at s _iProfile when inner, symbolic distance is for just, size is that (x is y) apart from s _iThe shortest Euclidean distance of profile; When (x y) just in time is positioned at s _iProfile when outside, symbolic distance is for negative, size is that (x is y) apart from s _iThe shortest Euclidean distance of profile.

D ((x, y), s _i) (x is y) to shape s for any point in the presentation video two dimensional surface _iThe shortest Euclidean distance of profile.

It should be noted that we have carried out alignment operation to the profile among shape in the template and the prominent edge figure in advance, make algorithm have rotation, translation invariance.

Said step 5,6 is realized by following steps;

(1) combine the gray scale geological information of prominent edge figure and the shape information that template provides, constitute energy function,

E＝E _snake+E _shape (12)

Wherein, E _SnakeSnake model energy function for traditional is defined as

$E_{\mathrm{snake}}={\∫}_0^{L\left(C\right)}g\left(\right|\▿I\left[C\left(p\right)\right]\left|\right)\mathrm{dp}---\left(13\right)$

The arc length of L in the formula (C) expression closed curve C; C (p) is the parametric representation form of closed curve, and when p got different values, C (p) was the coordinate of point under two-dimensional coordinate system different on the closed curve;

The shade of gray value at presentation video certain some place on closed curve; G () is a positive monotone decreasing and the zero function that levels off to, and can get usually $g(\▿I)=\frac{1}{1+\left|\right|\▿G\⊗I\left|\right|},$ (I is an original image, and G is a Gaussian function,

The expression convolution algorithm) is used for controlling level collection curve and moves, and finally stop near the edge E towards edge of image _ShapeBe defined as

$E_{\mathrm{shape}}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\γ}}_i{\mathrm{SD}}^2(C,C_i)---\left(14\right)$

C in the formula ₁, C ₂..., C _mProfile for m template providing; SD defines like (10) formula; γ _iBe weights, satisfy γ _i>=0, ∑ γ _i=1, concrete computing method are: certain template contours C _iWith the profile C among the prominent edge figure ₀Distance,

Divided by prominent edge figure and all template contours apart from sum

${\mathrm{\&gamma;}}_i=\frac{{\mathrm{SD}}^2({\mathrm{<figure-callout\; id="0"\; label="C"\; filenames="H2009100880689E00022.png"\; state="\{\{state\}\}">C</figure-callout>}}_0,C_i)}{\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{SD}}^2({\mathrm{<figure-callout\; id="0"\; label="C"\; filenames="H2009100880689E00022.png"\; state="\{\{state\}\}">C</figure-callout>}}_0,C_i)}---\left(15\right)$

(2) minimization of energy function: the rectangle corresponding with image boundary is the first closure profile, carries out curve evolvement along the negative gradient direction of energy function, promptly carries out the gradient decline iteration of energy function;

$\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}=-\frac{\&PartialD;E\left(\mathrm{\&phi;}\right)}{\&PartialD;\mathrm{\&phi;}}---\left(16\right)$

$\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}=g(\&dtri;I)|\&dtri;\mathrm{\&phi;}|\left(\mathrm{div}(\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}+v)\right)-\mathrm{\&beta;}\&dtri;E_{\mathrm{shape}}---\left(17\right)$

$\&dtri;E_{\mathrm{shape}}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}(\frac{\mathrm{\&phi;}-{\mathrm{\&phi;}}_i}{\left|\right|\mathrm{\&phi;}-{\mathrm{\&phi;}}_i\left|\right|}-\mathrm{div}\left(\frac{\&dtri;\left(\mathrm{\&phi;}{-\mathrm{\&phi;}}_i\right)}{\left|\right|\&dtri;\left(\mathrm{\&phi;}{-\mathrm{\&phi;}}_i\right)\left|\right|}\right))---\left(18\right)$

Wherein, v is a constant; Div representes divergence; β representes to be used for regulating the constant of the effect power of prior shape in energy function.Here φ is an imbedding function, and the zero level collection of imbedding function is exactly the target closed contour C that we will extract, promptly C={ (x, y) | φ (x, y, t)=0};

The initial value of said imbedding function is the symbolic distance function of first closure profile, and above-mentioned formula (11) is defined as according to formula (17) t evolution in time;

(3) if the distance between adjacent twice profile corresponding shape less than threshold epsilon, then stops iteration, the closed contour that obtain this moment is the final goal profile.

Beneficial effect of the present invention: proposed a kind of new target closed contour extraction algorithm.Through with the experiment confirm of this algorithm application in the multiclass natural image; This algorithm is a kind of effective objective contour method for distilling; Noisy background and partial occlusion are had robustness, thereby make this algorithm be applicable to multiple visual task, like target detection, tracking and identification etc.

Description of drawings

Fig. 1 is for extracting the process flow diagram of target closed contour.

The human primary visual cortex of Fig. 2 (a) has the cell receptive field figure of odd symmetry and even symmetry character;

The inhibiting effect synoptic diagram of the human primary visual cortex non-classical receptive field of Fig. 2 (b);

The prominent edge figure that Fig. 3 uses the step (1) of this algorithm to obtain;

(a) original image; (b) experimental result of the present invention; (c) experimental result of Canny edge extracting operator;

Fig. 4 template exemplary plot;

(a) template of in embodiment 1, using; (b) template of in embodiment 2, using;

The result of the similarity of the profile among shape in Fig. 5 calculation template and the prominent edge figure;

(a) the minimum situation (promptly the most similar) of distance; (b) distance situation placed in the middle; (c) the maximum situation (promptly least similar) of distance;

Fig. 6 uses the exemplary plot of the iterative process of this algorithm;

Fig. 7 is under noisy background and partial occlusion, and the present invention extracts the result of the accurate closed contour of target;

(a1) original image; (b1) experimental result of the present invention;

(a2) original image; (b2) experimental result of the present invention;

(a3) original image; (b3) experimental result of the present invention.

Embodiment

The present invention provides a kind of method of the extraction target closed contour based on shape prior, below in conjunction with accompanying drawing and embodiment invention is further specified.

Shown in Figure 1 for extracting the process flow diagram of target closed contour, the method for extracting target closed contour comprises following steps;

Step 1 is utilized the imageing sensor input picture;

Step 2 to the image of step 1 input, is set up the bank of filters p of Simulation of Complex cell ₁(f ₁, θ ₁)～p _n(f _n, θ _n), and image carried out Filtering Processing, obtain the edge;

Step 3 is set up the texture inhibition function r that simulates non-classical receptive field ₁(f ₁, θ ₁)～r _n(f _n, θ _n); (shown in the inhibiting effect synoptic diagram of cell receptive field figure and the human primary visual cortex non-classical receptive field of Fig. 2 (b) that the human primary visual cortex of Fig. 2 (a) has odd symmetry and even symmetry character);

Step 4 suppresses both synergy c to the edge extracting of step 2 Simulation of Complex cell and the texture of step 3 simulation non-classical receptive field ₁(f ₁, θ ₁)～c _n(f _n, θ _n) maximize processing, obtain prominent edge figure;

Step 5, the step of the similarity of shape among shape among calculation template 1～template m and the prominent edge figure;

Step 6, the step of the accurate closed contour of extraction target.

Experimental situation is PentiumIV2.4GHz CPU, the computing machine of 1G internal memory, and native system practical standard C++ and Matlab realize.The image library of using is the Weizmann image library, Cremers image library and INRIA PersonDataset image library.

The closed contour of embodiment 1 horse extracts

At first, obtain the prominent edge figure of input picture.Its course of work is described below:

1. the complex cell in the simulating human primary visual cortex from input picture, extract have CF and towards the mode of edge line segment, image I is carried out filtering by the Gabor bank of filters of one group of odd symmetry and even symmetry.Fig. 2 (a) has provided the cell receptive field figure that human primary visual cortex has odd symmetry and even symmetry character respectively, and the two-dimensional Gabor function at the image in spatial domain similarly.

2. the inhibiting effect in the simulating human primary visual cortex is with the difference of gaussian wave filter of normalization

Response p with complex cell _{F, θ}Convolution represent.Fig. 2 (b) has provided the inhibiting effect synoptic diagram of the classical receptive field of human primary visual cortex.

3. the edge extracting that extracts at complex cell adds inhibiting effect on, and when Gabor wave filter θ get different towards the time, respectively respond c from correspondence _{F, θ}In get maximum as prominent edge.Fig. 3 (a), Fig. 3 (b), Fig. 3 (c) have described the inhibiting prominent edge extraction effect of this consideration to noise and texture edge, as a comparison, give the result of Canny edge extracting operator commonly used.Can find out that the result of Canny edge extracting operator comprises many chickenshits and texture edge, and technology of the present invention has suppressed the texture edge of most of ground unrest edge and target internal effectively, has mainly kept significant profile.Can see also simultaneously that at the very approaching profile of locating of the contrast of prospect and background, use should technology also can't extract.

This just needs the curve evolvement technology based on template of back to obtain complete closed contour.

Secondly, the coarse profile among the calculating prominent edge figure and the similarity of each template, the weights of each shape of template when utilizing shape prior as next step.Fig. 4 (a) has provided 5 horse templates that this example is used, and Fig. 4 (b) has provided 6 people's templates that this example is used, and all is binary pictures.

The result of the similarity of the profile among shape in Fig. 5 calculation template and the prominent edge figure; Provided result according to the similarity of profile among the distance measurement prominent edge figure of formula (10) definition and template.Wherein

(a) the minimum situation (promptly the most similar) of distance; (b) distance situation placed in the middle; (c) the maximum situation (promptly least similar) two of distance

Can find out with prominent edge figure in contour shape approaching, the distance less; Have than large deformation with contour shape among the prominent edge figure, distance is just big.

Once more,, utilize, extract the accurate closed contour of target based on differential geometric curve evolvement technology according to the shape prior that template provides.Its course of work is described below:

1. the energy function that combines classical snake model and shape of template priori is set, and obtains the gradient expression-form of energy function;

2. be the first closure profile with the corresponding rectangle of image boundary, carry out iteration along the negative gradient direction of energy function;

3. the distance between adjacent twice profile corresponding shape stops iteration less than threshold epsilon, and the closed contour that obtain this moment is the final goal profile.

Fig. 6 is in the iterative process of energy function, 4 different middle corresponding constantly profile diagrams, and visible closed contour is shunk towards the border of target by initial rectangle gradually.

The objective contour that obtains when Fig. 7 is iteration stopping, wherein, (a1) original image; (b1) experimental result of the present invention;

(a2) original image; (b2) experimental result of the present invention; (a3) original image; (b3) experimental result of the present invention.Can find out that from this result the present invention has robustness to illumination, complex background and partial occlusion, can not produce the profile breakage problem, and when the target in the image and shape of template have than large deformation, also can produce profile accurately.

Embodiment 2 pedestrians' closed contour extracts

The unique difference part of the operation of the present embodiment and first embodiment is: when the coarse profile in calculating prominent edge figure and the similarity of each template; The size different to template taking; Get 6 sizes from equally spaced between full-sized 0.5 times to 2 times of the template; Calculate the similarity of different scaled lower bolster and prominent edge figure respectively, make similarity the highest (being that distance is minimum between shape) be of a size of matching size, follow-up step is all carried out under this matching size.This is that such pre-service is for making the present invention have the convergent-divergent unchangeability because pedestrian in the template and the pedestrian's size in the pending image have difference.

The information processing mechanism of simulating human vision; Merge shape prior; And utilize based on differential geometric curve evolvement technology, a kind of new target closed contour extraction algorithm VISDG (Human Vision InspiredIntegration of Shape Prior and Differential Geometry) has been proposed.The present invention has following outstanding characteristics: the first, and the VISDG algorithm necessarily can obtain complete target closed contour, and other contour extraction methods (such as the method based on cluster) are had to discrete profile fragment usually.Second; When image background is in a mess; The VISDG algorithm still can extract objective contour exactly, and this is to suppress texture among the human visual system with noise, only keep the mechanism of remarkable structure because we have simulated on the one hand, has introduced shape prior owing to us on the other hand.The 3rd, the VISDG algorithm has the robustness to partial occlusion.The 4th, the VISDG algorithm is not subject to illumination effect, even light causes that certain segment boundary of target is fuzzy a little less than excessively, still can extract the accurate target profile with the VISDG algorithm.The 5th, a small amount of template of a certain classification object only is provided, just can carry out profile to this type objects of different attitudes and extract, promptly the VISDG algorithm have to a certain degree to elastically-deformable unchangeability.

The present invention also exists and treats improvements: only considered to have only the situation of a well-marked target in the image at present, and do not considered that a plurality of targets are arranged in the image, need carry out the situation that profile separately extracts.

These instances only are used to the present invention is described and are not used in and limit usable range of the present invention; After having read teachings of the present invention; Those skilled in the art can do various changes and modification to this invention, and these equivalent form of values belong to the application's appended claims institute restricted portion equally.

Claims (1)

1. the method based on the extraction target closed contour of shape prior is characterized in that, comprises following steps;

Step 1 is utilized the imageing sensor input picture;

Step 2 to the image of step 1 input, is set up the bank of filters of Simulation of Complex cell, promptly by the Gabor bank of filters of one group of odd symmetry and even symmetry input picture is carried out filtering and obtains p ₁(f ₁, θ ₁)～p _n(f _n, θ _n), this is similar to the operation that the complex cell in the human primary visual cortex is taked input picture, purpose be from input picture, extract have CF and towards the edge line segment:

$p_{f,\mathrm{\&theta;}}=\sqrt{{\left(p_{f,\mathrm{\&theta;}}^e\right)}^2+{\left(p_{f,\mathrm{\&theta;}}^o\right)}^2}---\left(1\right)$

$p_{f,\mathrm{\&theta;}}^e=I(x,y)\&CircleTimes;h^e(x,y,\mathrm{\&theta;},f)---\left(2\right)$

$p_{f,\mathrm{\&theta;}}^o=I(x,y)\&CircleTimes;h^o(x,y,\mathrm{\&theta;},f)---\left(3\right)$

$h^e(x,y,\mathrm{\&theta;},f)=\exp (-\frac{1}{2}(\frac{x^{\&prime;2}}{{{\mathrm{\&sigma;}}_x}^2}+\frac{y^{\&prime;2}}{{{\mathrm{\&sigma;}}_y}^2}))\cos \left(2\mathrm{\&pi;f}{x}^{\&prime;}\right)---\left(4\right)$

$h^o(x,y,\mathrm{\&theta;},f)=\exp (-\frac{1}{2}(\frac{x^{\&prime;2}}{{{\mathrm{\&sigma;}}_x}^2}+\frac{y^{\&prime;2}}{{{\mathrm{\&sigma;}}_y}^2}))\sin \left(2\mathrm{\&pi;}{\mathrm{fx}}^{\&prime;}\right)---\left(5\right)$

Wherein x '=xcos θ+ysin θ y '=-xsin θ+ycos θ,

I (x, y) expression one width of cloth be of a size of M * N image in (x, the monochrome information of y) locating; p ^eThe Gabor wave filter h of expression even symmetry ^eTo the result of image filtering, p ^oRepresent odd symmetric Gabor wave filter h ^oTo the result of image filtering, p _{F, θ}The energy summation of representing these two kinds of filter filtering results, the response of corresponding complex cell;

The expression convolution algorithm, exp () expression is the exponent arithmetic at the end with e; F and θ represent respectively wave filter frequency and towards; σ _x ²And σ _y ²Represent the width of wave filter on x direction and y direction respectively; p ₁(f ₁, θ ₁)～p _n(f _n, θ _n) represent by a class frequency to be f ₁..., f _n, be oriented θ ₁..., θ _nThe bank of filters result that input picture carried out filtering;

Step 3 is set up the texture inhibition function r that simulates non-classical receptive field ₁(f ₁, θ ₁)～r _n(f _n, θ _n); Said texture inhibition function r ₁(f ₁, θ ₁)～r _n(f _n, θ _n) be established as and consider to connect the inhibiting effect that produces through side between the cell, calculating the energy that suppresses to produce is r _{F, θ}, purpose is to suppress noise edge and mixed and disorderly texture edge, thereby only keeps the edge of remarkable structure, this be the priority processing that in the long-term evolution process, forms of human visual system significantly or information of interest, ignore the strategy of redundant information:

$r_{f,\mathrm{\&theta;}}=p_{f,\mathrm{\&theta;}}\&CircleTimes;\widetilde{\mathrm{\&omega;}}---\left(6\right)$

$\mathrm{\&omega;}(x,y)=\frac{1}{4\mathrm{\&pi;}{\left(4\mathrm{\&sigma;}\right)}^2}\exp (-\frac{x^2+y^2}{4{\left(4\mathrm{\&sigma;}\right)}^2})-\frac{1}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\&sigma;}}^2})---\left(7\right)$

$\widetilde{\mathrm{\&omega;}}=\frac{\mathrm{\&omega;}}{\left|\right|\mathrm{\&omega;}\left|\right|},$

In the formula, σ representes the bandwidth of wave filter, || || expression norm, r _{F, θ}The expression inhibiting effect is with the difference of gaussian wave filter of normalization

Response p with complex cell _{F, θ}Convolution represent, be the non-classical receptive field pattern of simulation primary visual cortex cell;

Step 4 obtains prominent edge figure, comprising: from the edge that complex cell extracts, get rid of repressed edge earlier, promptly obtain the edge that keeps:

c _f，θ＝[p _f，θ-αr _f，θ] ₊ (8)

Wherein α is a constant, is used for regulating inhibiting power, [] ₊For getting non-negative operation, nonnegative number remains unchanged after operating through this, and negative is output as zero after operating through this;

Then, when θ get different towards the time, from each c of correspondence _{F, θ}In get maximum as prominent edge:

sc _f＝max{c _f，θ|θ＝θ ₁，θ ₂，…，θ _n} (9)

Step 5; Template set { template 1～template m} for the target to be detected that provides; Calculate the similarity of the contour shape among shape and the prominent edge figure that is obtained by step 4 in each template through the definition of the distance between shape and the shape, wherein the distance definition between the shape is following: suppose s _iAnd s _jBe any two shapes, SD (s _i, s _j) represent the distance between them

$\mathrm{SD}(s_i,s_j)=\sqrt{{\left|\right|\mathrm{\&phi;}\left(s_i\right)-\mathrm{\&phi;}\left(s_j\right)\left|\right|}^2+{\left|\right|\&dtri;\mathrm{\&phi;}\left(s_i\right)-\&dtri;\mathrm{\&phi;}\left(s_j\right)\left|\right|}^2}---\left(10\right)$

Wherein, φ (s _i) be shape s _iThe symbolic distance function, (x y) just in time is positioned at s to any point in the image two dimensional surface _iProfile on the time, symbolic distance is 0; When (x y) is positioned at s _iProfile when inner, symbolic distance is for just, size is that (x is y) apart from s _iThe shortest Euclidean distance of profile; When (x y) just in time is positioned at s _iProfile when outside, symbolic distance is for negative, size is that (x is y) apart from s _iThe shortest Euclidean distance of profile,

D ((x, y), s _i) (x is y) to shape s for any point in the presentation video two dimensional surface _iThe Euclidean distance of profile; It should be noted that in advance and will carry out alignment operation, make algorithm have rotation, translation invariance the profile among shape in the template of target to be detected and the prominent edge figure;

Step 6 is obtained the closed contour of target through following fine division step, comprises

(1) combine the gray scale geological information of prominent edge figure and the shape information that template provides, constitute energy function,

E＝E _snake+E _shape (12)

Wherein, E _SnakeSnake model energy function for traditional is defined as

$E_{\mathrm{snake}}={\&Integral;}_0^{L\left(C\right)}g\left(\right|\&dtri;I\left[C\left(p\right)\right]\left|\right)\mathrm{dp}---\left(13\right)$

The arc length of L in the formula (C) expression closed curve C; C (p) is the parametric representation form of closed curve, and when p got different values, C (p) was the coordinate of point under two-dimensional coordinate system different on the closed curve;

The shade of gray value at presentation video certain some place on closed curve; G () is a positive monotone decreasing and the zero function that levels off to, and gets

I is an original image, and G is a Gaussian function,

The expression convolution algorithm; Be used for controlling level collection curve and move, and finally stop near the edge E towards edge of image _ShapeBe defined as

$E_{\mathrm{shape}}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_i{\mathrm{SD}}^2(C,C_i)---\left(14\right)$

C in the formula ₁, C ₂..., C _mProfile for m template of the target to be detected that provides; SD defines like (10) formula; γ _iBe weights, satisfy γ _i>=0, ∑ γ _i=1, concrete computing method are: certain template contours C _iWith the profile C among the prominent edge figure ₀Distance, divided by prominent edge figure and all template contours apart from sum,

${\mathrm{\&gamma;}}_i=\frac{{\mathrm{SD}}^2(C_0,C_i)}{\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{SD}}^2(C_0,C_i)}---\left(15\right)$

(2) minimization of energy function: the rectangle corresponding with image boundary is the first closure profile, carries out curve evolvement along the negative gradient direction of energy function, promptly carries out the gradient decline iteration of energy function;

$\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}=-\frac{\&PartialD;E\left(\mathrm{\&phi;}\right)}{\&PartialD;\mathrm{\&phi;}}---\left(16\right)$

$\frac{\&PartialD;\mathrm{\&phi;}}{\&PartialD;t}=g(\&dtri;I)|\&dtri;\mathrm{\&phi;}|(\mathrm{div}\left(\frac{\&dtri;\mathrm{\&phi;}}{|\&dtri;\mathrm{\&phi;}|}\right)+v)-\mathrm{\&beta;}\&dtri;E_{\mathrm{shape}}---\left(17\right)$

$\&dtri;E_{\mathrm{shape}}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}(\frac{{\mathrm{\&phi;}-\mathrm{\&phi;}}_i}{\left|\right|\mathrm{\&phi;}-{\mathrm{\&phi;}}_i\left|\right|}-\mathrm{div}\left(\frac{\&dtri;(\mathrm{\&phi;}-{\mathrm{\&phi;}}_i)}{\left|\right|\&dtri;(\mathrm{\&phi;}-{\mathrm{\&phi;}}_i)\left|\right|}\right))---\left(18\right)$

Wherein, v is a constant; Div representes divergence; β representes to be used for regulating the constant of the effect power of prior shape in energy function; φ is an imbedding function, and the zero level collection of imbedding function is exactly the target closed contour C that we will extract, promptly C={ (x, y) | φ (x, y, t)=0};

The initial value of said imbedding function is the symbolic distance function of first closure profile, defines like above-mentioned formula (11); Imbedding function is according to formula (17) t evolution in time;

(3) if the distance between adjacent twice profile corresponding shape less than threshold epsilon, then stops iteration, the closed contour that obtain this moment is the final goal profile.

Images