CN104766326A - Shape matching locating method and device based on yin-yang discrete point sampling model - Google Patents

Abstract

The invention discloses a shape matching locating method based on a yin-yang discrete point sampling model. The method comprises the steps that first, shape contour templates are extracted from yin discrete point sampling images or yang discrete point sampling images in sample images; second, the extracted shape contour templates are subjected to optimization, and the best shape contour template is obtained; and third, yin discrete point sampling images or yang discrete point sampling images of practical images are obtained and are matched with the corresponding best shape contour template, so that images are recognized. In the third step, matching is achieved through a matching ratio operator for computing density estimation of the yin discrete point sampling images or the yang discrete point sampling images of the practical images and the corresponding best shape contour template. The invention further discloses a corresponding shape matching locating device. According to the method and device, the shape contour templates are extracted from the yin discrete point sampling images or the yang discrete point sampling images in the sample images, matching of the detected images is achieved, a shape to be detected is recognized, and quick extracting and accurate distinguishing of line-face combination shapes under a complex background in current image recognition are achieved.

Description

A kind of form fit localization method based on negative and positive discrete point sampling model and device

Technical field

The present invention relates to images match identification field, particularly relate to the method and device of in a kind of digital picture, carrying out form fit location.

Background technology

Utilize order target area, boundary information realize identifying from complicated image and the form fit of localizing objects cognitive and distinguish the universal law of target owing to meeting most human eye, thus obtain image procossing, the large quantities of scholar of computer vision field study for a long time widely.Form fit model comprise scene graph in the middle of express, shape template and matching way three key elements.Detect in the middle of the application of target in hope fast by form fit, to the requirement mainly stability expressed in the middle of scene graph, namely the impact that should not be subject to illumination variation, speckle noise, trickle geometry deformation in shooting process is as far as possible expressed in this centre; Not only to the requirement of shape template mainly representative and separability, this template can represent this class target but also can will make a distinction with other class targets; To the rapidity that the requirement of matching way mainly calculates.

Form fit can be divided into the form fit based on edge and the form fit based on region two class from the middle expressive perspective of scene graph.Form fit based on edge often requires clear and definite point-to-point skirt response, and rim detection is a difficult categorised decision, to noise and illumination variation responsive.In order to strengthen the stability that outline map is expressed as centre, have method edge chips to be organized into groups, the curve separability after marshalling is stronger, has less potential coupling than isolated edge chips, this makes coupling relatively easily carry out, and can block and work under edge deletion condition.Other methods then by carry out form fit to avoid in the segmentation figure of over-segmentation a little with point skirt response, obtain to compact shape change and space change adaptive faculty.But these methods all have higher computation complexity usually, be difficult to the demand meeting real-time online detection.Coupling based on region does not need clear and definite skirt response, to the distortion of local shape, to block robustness stronger, but lack important shape details, complicated shape is difficult to stable extraction, expresses and exact matching, and its classification capacity is restricted in complex scene.

A kind of image analysis method based on negative and positive discrete point sampling model is disclosed in applicant patent CN201110349510A formerly, it is according to the type of target to be detected in image, size and degree of noise interference select sample operator, sample radius and neighborhood gray uniformization pass judgment on threshold value, and then utilize above-mentioned sample operator, sample radius and neighborhood gray uniformization pass judgment on threshold value, then calculate by carrying out sampling to image, obtain corresponding cloudy sample graph and positive sample graph, finally in described cloudy sample graph and positive sample graph, discrete point grouping method is utilized to detect corresponding target.The method considers the advantage of based target region and based target border two class methods simultaneously, the boundary information of target in target image is obtained by cloudy sample graph, the area information of target in target image is obtained by positive sample graph, yin, yang sample graph reflects the transient process of objective area in image to object edge jointly, sample radius can switch according to the change of target size simultaneously, computing velocity is significantly improved, also has stronger adaptability to the change in size of target.

But, in the method, underuse the priori of known form, the deficiency that range of application is limited to.Such as discrete point grouping method can only be tieed up according to N the discrete point that connectivity pair that peripheral direction contribution degree describes has certain connected relation and organize into groups, if the connectedness of discrete point is interrupted due to the interference of complex background, discrete point marshalling just cannot go on, usually organize into groups result because the interference of complex background can not extract the shape expected simultaneously, extract too many unwanted center line in other words, cause its line in images match under complex background, face combined shaped cannot rapid extraction and accurately distinguishing, make images match discrimination not high.

Summary of the invention

For above defect or the Improvement requirement of prior art, the invention provides a kind of form fit localization method based on negative and positive discrete point sampling model, it by extracting shape contour mould from the cloudy discrete point sample graph or positive discrete point sample graph of sample image, realize the coupling of detected image and identify shape to be detected, solve current image recognition detect in line under complex background, face combined shaped rapid extraction and accurately distinguish problem.

For achieving the above object, according to one aspect of the present invention, provide a kind of form fit localization method based on negative and positive discrete point sampling model, it realizes images match by the shape contour mould extracted in cloudy discrete point sample graph or positive discrete point sample graph, it is characterized in that, the method comprises:

The first step of shape contour mould is extracted from the cloudy discrete point sample graph sample image or positive discrete point sample graph;

Carry out preferably to the shape contour mould extracted, obtain the second step of optimum shape contour mould: and

Obtain the cloudy discrete point sample graph of real image or positive discrete point sample graph, it is mated with corresponding optimum shape contour mould, thus the third step of recognition image;

Wherein, in described third step, obtain with the matching rate operator of the density Estimation of corresponding optimum shape contour mould especially by the cloudy discrete point sample graph or positive discrete point sample graph that calculate described real image mating.

As improvement of the present invention, cloudy discrete point sample graph is utilized to carry out the matching rate operator of described density Estimation:

$\mathrm{\λ}=\frac{N}{M}=\frac{\underset{k=1}{\overset{M}{\mathrm{\Σ}}}S\left(p_k\right)}{M}$

In formula, M is summation of counting in shape contour mould, and N is the summation of counting that on shape contour mould, each point has response relation in discrete point sample graph banded zone, p _kbe current point on shape contour mould, arbitrary curve is by the response relation S (P of certain discrete point _k)

$S\left(p_k\right)=\left\{\begin{array}{cc}1& I\left(p_k\right)=1\mathrm{or}{C}_8\left(p_k\right)\&GreaterEqual;2\\ 0& C_8\left(p_k\right)<2\end{array}\right.$

As improvement of the present invention, the matching rate operator utilizing positive discrete point sample graph to carry out described density Estimation is:

$\mathrm{\&gamma;}=\frac{N}{M}\&times;r^2$

In formula, M is summation of counting in region template, and N is region template discrete point summation in overlay area in discrete point sample graph, and r is the sample radius of negative and positive discrete point sampling model.

As improvement of the present invention, described matching rate operator can comprehensive positive discrete point sample graph matching rate operator and cloudy discrete point sample graph matching rate operator obtain.

As improvement of the present invention, the structure conspicuousness that the described shape contour mould to extracting carries out preferably by calculating shape profile template matches feature realizes, and wherein, the structure conspicuousness of described shape profile template matches feature is specifically calculated as follows:

If there is n class target, and the effect of feature t is by class ω _imake a distinction with other n-1 classes, then define this feature t for class ω _iarchitectural feature significance measure

$S_i^t=\frac{1}{n-1}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{S}_{\mathrm{ij}}^t,j\&NotEqual;i$

Wherein, $S_{\mathrm{ij}}^t=1-e^{-a{\left(\frac{{(m_i-m_j)}^2}{{\mathrm{\&sigma;}}_i^2+{\mathrm{\&sigma;}}_j^2+c}\right)}^b},0<a<\mathrm{1,0}<b<1$

In formula, m _iand m _jfor class ω _iwith class ω _jthe average of corresponding templates matching characteristic t, with it is its variance.Formula (3-5) shows that between class, difference is larger, and class interpolation is less, namely larger, feature t is more remarkable, more can by class ω _iwith class ω _jmake a distinction.In formula the effect of natural logarithm e be by be adjusted to the value between a 0-1, so that compare between the feature of different span.A is proportional control factor, and b is index adjustment factor, and both cooperations can make difference calculated value adjust to be convenient to distinguish degree, c is a small constant, is 0 for avoiding denominator.

As improvement of the present invention, there is convergent-divergent, rotation geometry is out of shape between shape to be extracted and described preferred shape contour mould, then need to carry out region conversion to shape contour mould, affine deformation Matrix Formula is:

$\left[\begin{array}{ccc}\mathrm{\&alpha;}& \mathrm{\&beta;}& (1-\mathrm{\&alpha;})\&times;\mathrm{center}.x-\mathrm{\&beta;}\&times;\mathrm{cebter}.y\\ -\mathrm{\&beta;}& \mathrm{\&alpha;}& \mathrm{\&beta;}\&times;\mathrm{center}.x+(1-\mathrm{\&alpha;})\&times;\mathrm{center}.y\end{array}\right]$

Wherein α=scale × cos (angle), β=scale × sin (angle), the scale coefficient of the corresponding affine deformation of scale, the coefficient of rotary of the corresponding affine deformation of angle.Center.x and center.y represents horizontal ordinate and the ordinate of convergent-divergent or rotation center respectively.

As improvement of the present invention, under can using estimating target deformation condition, directly can determine deformation formula and then directly obtain corresponding contour mould, that is: first according to the priori of shape to be extracted, select from the center line extracted most possibly belong to shape to be extracted one or several center line as key central line, the size possible to shape contour mould, direction, position is estimated, then in cloudy discrete point sample graph corresponding key central line discrete point banded zone in coupling checking this shape contour mould estimated whether appropriate.

According to another aspect of the present invention, a kind of form fit locating device based on negative and positive discrete point sampling model is provided, it realizes images match by the shape contour mould extracted in cloudy discrete point sample graph or positive discrete point sample graph, and it is characterized in that, this device comprises:

First module, it for extracting shape contour mould from the cloudy discrete point sample graph in sample image or positive discrete point sample graph;

Second module, it, for carrying out preferably to the shape contour mould extracted, obtains optimum shape contour mould: and

3rd module, it, for obtaining the cloudy discrete point sample graph of real image or positive discrete point sample graph, mates with corresponding optimum shape contour mould by it, thus recognition image;

Wherein, in described 3rd module, obtain with the matching rate operator of the density Estimation of corresponding optimum shape contour mould especially by the cloudy discrete point sample graph or positive discrete point sample graph that calculate described real image mating.

As improvement of the present invention, cloudy discrete point sample graph is utilized to carry out the matching rate operator of described density Estimation:

$\mathrm{\&lambda;}=\frac{N}{M}=\frac{\underset{k=1}{\overset{M}{\mathrm{\&Sigma;}}}S\left(p_k\right)}{M}$

In formula, M is summation of counting in shape contour mould, and N is the summation of counting that on shape contour mould, each point has response relation in discrete point sample graph banded zone, p _kbe current point on shape contour mould, arbitrary curve is by the response relation S (P of certain discrete point _k)

$S\left(p_k\right)=\left\{\begin{array}{cc}1& I\left(p_k\right)=1\mathrm{or}{C}_8\left(p_k\right)\&GreaterEqual;2\\ 0& C_8\left(p_k\right)<2\end{array}\right.$

As improvement of the present invention, the matching rate operator utilizing positive discrete point sample graph to carry out described density Estimation is:

$\mathrm{\&gamma;}=\frac{N}{M}\&times;r^2$

In formula, M is summation of counting in region template, and N is region template discrete point summation in overlay area in discrete point sample graph, and r is the sample radius of negative and positive discrete point sampling model, is multiplied by r ²in order to the pixel caused because of discrete sampling vacancy factor is taken into account.

As improvement of the present invention, the structure conspicuousness that the described shape contour mould to extracting carries out preferably by calculating shape profile template matches feature realizes, and wherein, the structure conspicuousness of described shape profile template matches feature is specifically calculated as follows:

If there is n class target, and the effect of feature t is by class ω _imake a distinction with other n-1 classes, then define this feature t for class ω _iarchitectural feature significance measure

$S_i^t=\frac{1}{n-1}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{S}_{\mathrm{ij}}^t,j\&NotEqual;i$

Wherein, in formula, m _iand m _jfor class ω _iwith class ω _jthe average of corresponding templates matching characteristic t, with it is its variance.Formula (3-5) shows that between class, difference is larger, and class interpolation is less, namely larger, feature t is more remarkable, more can by class ω _iwith class ω _jmake a distinction.In formula the effect of natural logarithm e be by be adjusted to the value between a 0-1, so that compare between the feature of different span.A is proportional control factor, and b is index adjustment factor, and both cooperations can make difference calculated value adjust to be convenient to distinguish degree, c is constant.

As improvement of the present invention, there is convergent-divergent, rotation geometry is out of shape between shape to be extracted and described preferred shape contour mould, then need to carry out region conversion to shape contour mould, affine deformation Matrix Formula is:

$\left[\begin{array}{ccc}\mathrm{\&alpha;}& \mathrm{\&beta;}& (1-\mathrm{\&alpha;})\&times;\mathrm{center}.x-\mathrm{\&beta;}\&times;\mathrm{cebter}.y\\ -\mathrm{\&beta;}& \mathrm{\&alpha;}& \mathrm{\&beta;}\&times;\mathrm{center}.x+(1-\mathrm{\&alpha;})\&times;\mathrm{center}.y\end{array}\right]$

Wherein α=scale × cos (angle), β=scale × sin (angle), the scale coefficient of the corresponding affine deformation of scale, the coefficient of rotary of the corresponding affine deformation of angle, center.x and center.y represents horizontal ordinate and the ordinate of convergent-divergent or rotation center respectively.

As improvement of the present invention, under can using estimating target deformation condition, directly can determine deformation formula and then directly obtain corresponding contour mould, that is: first according to the priori of shape to be extracted, select from the center line extracted most possibly belong to shape to be extracted one or several center line as key central line, the size possible to shape contour mould, direction, position is estimated, then in cloudy discrete point sample graph corresponding key central line discrete point banded zone in coupling checking this shape contour mould estimated whether appropriate.

In the present invention, the introducing of contour mould can refuse rapidly unwanted center line, to noise with fuzzyly have very strong immunity; Discrete point sample graph is extremely suitable for carrying out density Estimation to contour mould, as long as there is the banded zone that a slice discrete point aggregates in sample graph, certain contour mould can be admitted inside it, just this contour mould is contained with very large probability in former figure, this contour mould itself does not need to have continuity, can be the independent assortment of some key central lines, thus greatly extend the scope of application of carrying out graphical analysis based on negative and positive discrete point sampling model.

In general, the above technical scheme conceived by the present invention compared with prior art, has following beneficial effect:

(1) according to the concrete condition of application, in negative and positive discrete point sample graph, corresponding shape contour mould can be obtained by the method for Freehandhand-drawing or discrete point marshalling voluntarily, simple and practical.

(2) calculated by the conspicuousness of contour mould, can set up and a set ofly select the practical step being more conducive to carrying out the shape profile form assembly that shape extracting and similar shape are distinguished, user can by repeatedly drawing or generating different shape contour moulds and calculate teaching display stand optimization by conspicuousness.

(3) calculated by the matching rate of the matching rate of density Estimation, the shape that can fast, robustly combine line, face under complex background is carried out extract real-time and accurately distinguishes

(4) estimated and affine deformation by target distortion, can fast processing when there is convergent-divergent between shape to be extracted and described preferred shape contour mould, situation that rotation geometry is out of shape.

Accompanying drawing explanation

Fig. 1 is the four kinds of contour moulds utilizing certain wheel arrangement type of manual drawing to adopt in the method according to the embodiment of the present invention;

Fig. 2 is with the contour mould for distinguishing five class wheel arrangements that structure conspicuousness is picked out in the method according to the embodiment of the present invention;

Fig. 3 utilizes the wheel arrangement five class contour mould in Fig. 2 to carry out matching rate λ _ithe result calculated;

Fig. 4 be in the method according to the embodiment of the present invention under various noise based on contour mould estimate wheel alignment result.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

A kind of shape profile matching process based on negative and positive discrete point sample graph constructed by the present embodiment, comprise and extract shape contour mould from the cloudy discrete point sample graph or kation discrete point sample graph of sample image, the template extracted is carried out preferably and utilizes the shape contour mould preferably obtained to mate these processes.

Particularly, in cloudy discrete point sample graph, shape contour mould is mated, calculate the matching rate between them by density Estimation.The introducing of shape contour mould can refuse rapidly unwanted profile and elongated curve, to noise with fuzzyly have very strong immunity; And cloudy discrete point sample graph is extremely suitable for carrying out density Estimation to shape contour mould, as long as there is the banded zone that a slice discrete point aggregates in sample graph, certain shape contour mould can be admitted inside it, in former figure, just contain this shape contour mould with very large probability.

In the present embodiment, in cloudy discrete point sample graph to the matching rate operator that shape contour mould carries out density Estimation be

$\mathrm{\&lambda;}=\frac{N}{M}=\frac{\underset{k=1}{\overset{M}{\mathrm{\&Sigma;}}}S\left(p_k\right)}{M}---\left(1\right)$

In formula (1), M is summation of counting in shape contour mould, and N is the summation of counting that on shape contour mould, each point has response relation in discrete point sample graph banded zone, p _kbe current point on shape contour mould, arbitrary curve is by the response relation S (P of certain discrete point _k) define and see formula (2).

$S\left(p_k\right)=\left\{\begin{array}{cc}1& I\left(p_k\right)=1\mathrm{or}{C}_8\left(p_k\right)\&GreaterEqual;2\\ 0& C_8\left(p_k\right)<2\end{array}\right.---\left(2\right)$

In formula, I (p _k) for record in cloudy discrete point sample graph with curve current point p _kwhether correspondence position exists discrete point, if there is discrete point, then this point is black pixel, and value is 1, if there is no discrete point, then this point is white pixel, and value is 0.I (p _kwith curve current point p in the cloudy discrete point sample graph of)=1 expression _kcorrespondence position is black pixel, and value is C ₈(p _krepresent in cloudy discrete point sample graph with p _kblack number of pixels around correspondence position on 8 UNICOM directions, C ₈(p _kwith p in the cloudy discrete point sample graph of)>=2 expression _karound correspondence position, there is the black pixel of two or more in 8 UNICOM directions, and both of these case is all identified as curve by a discrete point in cloudy discrete point sample graph, otherwise does not then pass through.Avoiding problems clear and definite point-to-point corresponding, have stronger resistivity to the geometry deformation in certain limit and noise.

The Algorithms T-cbmplexity mated shape contour mould in cloudy discrete point sample graph, far below the time complexity of conventional gray scale template matches and edge matching algorithm.

In the present embodiment, the shape contour mould extracted is carried out preferably, if there is not large disparity between shape to be matched and shape contour mould, and require find out in several similar shape closest to shape contour mould that time situation, namely how to realize the accurate differentiation of similar shape.The difficulty that outline reduces form fit is carried out in cloudy discrete point sample graph, but the separability of shape contour mould is had higher requirement, need to measure the structure conspicuousness of multiple shape profile template matches feature, thus pick out one group of representative and that separability is strong shape contour mould and carry out coupling and classify.

Wherein, representative this template of finger by force can tolerate the disparity between similar target as far as possible, and separability refers to that by force this template can distinguish inhomogeneous target as far as possible.Feature significance measurement criterion can be divided into structure conspicuousness and probability conspicuousness two kinds, structure conspicuousness using in class, between class distance function as criterion, little as far as possible by structure inter-object distance, and the large as far as possible criterion function of between class distance selects shape template; Probability conspicuousness, then using minimum error probability as criterion, is picked out on the whole, is statistically judged minimum shape template by accident.The advantage of structure conspicuousness only needs a small amount of sample just can learn, and shortcoming is that the accuracy rate of classification is high not as probability conspicuousness.When a small amount of sample only can be obtained, structure conspicuousness can be adopted to learn.Probability conspicuousness advantage based on minimum error probability is that accuracy rate is higher, and shortcoming needs a large amount of learning samples.

In this programme, regard each shape contour mould as an observer, each coupling output matching characteristic t, then the structure conspicuousness set up based on matching characteristic t selects the method for shape contour mould, by drawing in the discrete point sample graph of learning sample, defining and the structure conspicuousness calculating shape template selects the shape template that more can represent this class target and more other class targets can be made a distinction again, this process can be divided into following three steps to carry out:

(1) single template matches feature is measured for distinguishing the structure conspicuousness of (such as two class targets)

If there are two class target ω _iand ω _j, and the effect of template characteristic t is by class ω _iwith class ω _jmake a distinction, defined feature t is for class ω _iwith class ω _jarchitectural feature significance measure

$S_{\mathrm{ij}}^t=1-e^{-a{\left(\frac{{(m_i-m_j)}^2}{{\mathrm{\&sigma;}}_i^2+{\mathrm{\&sigma;}}_j^2+c}\right)}^b},0<a<\mathrm{1,0}<b<1---\left(3\right)$

In formula (3), m _iand m _jfor class ω _iwith class ω _jthe average of corresponding templates matching characteristic t, with it is its variance.Formula (3-5) shows that between class, difference is larger, and class interpolation is less, namely larger, feature t is more remarkable, more can by class ω _iwith class ω _jmake a distinction.In formula the effect of natural logarithm e be by be adjusted to the value between a 0-1, so that compare between the feature of different span.A is proportional control factor, and b is index adjustment factor, and both cooperations can make difference calculated value adjust to be convenient to distinguish degree, c is a small constant, is 0 for avoiding denominator.In the present embodiment, can calculate by experiment and preferably work as a=0.3, during b=0.7 calculated value to distinguish degree comparatively large, therefore preferably get a=0.3 in the present embodiment, b=0.7, c=0.0000001.

(2) single template characteristic is measured for distinguishing the structure conspicuousness of (such as multi-class targets)

If there is n class target, and the effect of feature t is by class ω _imake a distinction with other n-1 classes, then define this feature t for class ω _iarchitectural feature significance measure

$S_i^t=\frac{1}{n-1}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{S}_{\mathrm{ij}}^t,j\&NotEqual;i---\left(4\right)$

Formula (4) shows be worth larger, feature t is more remarkable, more can by class ω _imake a distinction with other n-1 classes.

(3) optimal Template combination is selected

Suppose there is multiple template, its matching characteristic t ₁, t ₂..., t _kall have class ω _iability with other n-1 classes make a distinction, then calculate each feature t respectively _ifor class ω _iarchitectural feature significance measure get

$S_i=\max \left(S_i^{t_k}\right)---\left(5\right)$

Characteristic of correspondence is as class ω _ibest features, is denoted as t ⁱ.If all adopt said method to select best features to each class target, a multiclass notable feature combination can be obtained:

T＝{t ¹,t ²,…t ⁿ} (6)

Preferably, if there is the geometry deformations such as larger convergent-divergent, rotation between shape to be extracted and the shape contour mould of specifying, also need to carry out region conversion to shape contour mould.

Wherein affine deformation matrix is as shown in formula (7):

$\left[\begin{array}{ccc}\mathrm{\&alpha;}& \mathrm{\&beta;}& (1-\mathrm{\&alpha;})\&times;\mathrm{center}.x-\mathrm{\&beta;}\&times;\mathrm{cebter}.y\\ -\mathrm{\&beta;}& \mathrm{\&alpha;}& \mathrm{\&beta;}\&times;\mathrm{center}.x+(1-\mathrm{\&alpha;})\&times;\mathrm{center}.y\end{array}\right]---\left(7\right)$

Wherein α=scale × cos (angle), β=scale × sin (angle), the scale coefficient of the corresponding affine deformation of scale, the coefficient of rotary of the corresponding affine deformation of angle, center.x and center.y represents horizontal ordinate and the ordinate of convergent-divergent or rotation center respectively.

Such as, in an embodiment preferably corresponding scale coefficient scale value is 1.1, dimension deformation number of times is 10, corresponding angle coefficient of angularity is 5, one direction rotational deformation number of times is 3, altogether can obtain 10 × (2 × 3+1)=70 shape contour mould, and corresponding 70 region templates.

After obtaining above-mentioned template, calculate their matching rates in discrete point sample graph one by one.Preferably the cloudy matching rate of shape contour mould in cloudy sample graph not only will be calculated in the present embodiment, also can consider the positive matching rate of region template in positive sample graph changed into by shape contour mould, preferably comprehensive two matching rates provide final coupling positioning result.

Wherein, the computing formula of positive matching rate is shown in formula (8)

$\mathrm{\&gamma;}=\frac{N}{M}\&times;r^2---\left(8\right)$

In formula (8), M is summation of counting in region template, and N is region template discrete point summation in overlay area in discrete point sample graph, and r is the sample radius of negative and positive discrete point sampling model, is multiplied by r ²in order to the pixel caused because of discrete sampling vacancy factor is taken into account.

In theory, if only there is affine deformation between shape to be extracted and the shape contour mould of specifying, then always can find by the method for exhaustion profile and region template that mate shape to be extracted most, but in actual applications, unlimited many time can not be provided to attempt all possibilities, if the direction at this moment having clue to indicate affine deformation most possible, then can greatly reduce the number of times and time that calculate profile and region template.

Based on above-mentioned thinking, the shape extracting method that Shape-based interpolation contour mould is estimated can be set up.The center line that a upper joint extracts just can play this effect.First according to the priori of shape to be extracted, select from the center line extracted most possibly belong to shape to be extracted one or several center line as key central line, the possible size of shape contour mould, direction, position are estimated, then in cloudy discrete point sample graph corresponding key central line discrete point banded zone in this shape contour mould estimated of coupling checking whether appropriate.

Form fit based on contour mould can realize the function of part classification and incipient fault zone location in TFDS, and key is the quality how improving contour mould, enables current type parts and other types to be made a distinction better.

The concrete grammar that the present embodiment is optimized for wheel arrangement pattern of descriptive parts.As shown in Figure 1, for a certain specific wheel unit type, first can prepare the contour mould of several candidate more, as Fig. 1 (a)-(d), then structure significant characteristics is utilized to measure, with matching rate λ for feature t, calculate the architectural feature conspicuousness of contour mould for this unit type of these candidates with formula (4) select the highest contour mould as shape template, thus improves the accuracy of part classification and incipient fault zone location.

In the present embodiment, if in the middle of four of Fig. 1 contour moulds, contour mould 1 the highest, wheel arrangement Class1 and other wheel arrangement type classification can be come.Above-mentioned data based on the every class 200 from Test Field random choose, 1000 wheel arrangement Class1-5 pictures altogether.

The highest contour mould of structure significance measure value is selected to each wheel arrangement type method all applied above that reality uses, finally can obtain one group of remarkable configuration form assembly, as shown in Figure 2.Set up a shape template storehouse based on this group contour mould combination, the matching rate λ of formula (1) calculating to each contour mould i is used successively to each width test pattern _i, select λ _iwheel arrangement type corresponding to the highest contour mould is the wheel arrangement type of present image.

Fig. 3 figure provides and utilizes the wheel arrangement five class contour mould in Fig. 2 to carry out matching rate λ _ithe result calculated, can find out that this 5 class Freehandhand-drawing template has good separability, directly can pass through more different λ _isize classify.

In the present embodiment, the storage format of preferred each contour mould is: starting point absolute coordinates (4Byte)+starting point bearing of trend (1Byte)+starting point development length (2Byte)+contour mould point relative coordinate sequence (4Byte × (M-1), M are summations of counting in contour mould).If carefully select starting point during drawing image and record its bearing of trend and development length as necessary condition, just a large amount of extraneous areas can be got rid of rapidly during coupling, quickening matching speed.

In addition, in the present embodiment, with the center line extracted as a reference, first according to the priori of shape to be extracted, select from the center line extracted most possibly belong to shape to be extracted one or several center line as key central line, the possible size of contour mould, direction, position are estimated, then in cloudy discrete point sample graph corresponding key central line discrete point banded zone in this contour mould estimated of coupling checking whether appropriate.The test problems that this solves from different size, position wheel in 5 boxcar fault detect websites is adopted in the present embodiment.

First the wheel alignment method estimated based on contour mould finds the curve belonging to wheel as key central line from detected center line, the curve belonging to right wheel should meet one of following three conditions: the circular curve of (1) radius between 200-400 pixel or partial arc curve; (2) rail straight line intersection corresponding to lower right side, bottom, the non-rectilinear that connects, meet; (3) the bottom right district of discrete point sample graph may belong to the center line of right wheel border banded zone.

The algorithm curve that according to condition (1)-(3) this three conditions of sequential search are corresponding, curve such in the central line pick-up figure of nearly all wheel image all can have 1-2 bar.After obtaining belonging to the key central line on wheel border, the circle of multiple radiuses between 200-400 pixel that all curvilinear structures unique points of key central line can be formed, in discrete point sample graph, calculate the number of each round bottom right semi arch by discrete point, the final wheel alignment result exported can be by the circular arc place circle of maximum discrete point.Fig. 4 gives five width and is disturbed on image and uses said method to carry out the result of wheel alignment.

Those skilled in the art will readily understand; the foregoing is only preferred embodiment of the present invention; not in order to limit the present invention, all any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within protection scope of the present invention.

Claims (10)

1., based on a form fit localization method for discrete point sampling model, it realizes images match by utilizing the shape contour mould extracted in discrete point sample graph, and it is characterized in that, the method comprises:

The first step of shape contour mould is extracted from the cloudy discrete point sample graph sample image;

Carry out preferably to the shape contour mould extracted, obtain the second step of optimum shape contour mould: and

Obtain the cloudy discrete point sample graph of real image and/or positive discrete point sample graph, it is mated with corresponding optimum shape contour mould, thus the third step of recognition image;

Wherein, in described third step, carry out mating obtaining with the matching rate operator of the density Estimation of corresponding optimum shape contour mould especially by the cloudy discrete point sample graph or positive discrete point sample graph that calculate described real image.

2. a kind of form fit localization method based on discrete point sampling model according to claim 1, wherein, utilizes cloudy discrete point sample graph to carry out the matching rate operator of described density Estimation:

$\mathrm{\&lambda;}=\frac{N}{M}=\frac{\underset{k=1}{\overset{M}{\mathrm{\&Sigma;}}}S\left(p_k\right)}{M}$

In formula, M is summation of counting in shape contour mould, and N is the summation of counting that on shape contour mould, each point has response relation in discrete point sample graph banded zone, p _kbe current point on shape contour mould, arbitrary curve by the response relation of this discrete point is $S\left(p_k\right)=\left\{\begin{array}{cccc}1& I\left(p_k\right)=1& \mathrm{or}& C_8\left(p_k\right)\&GreaterEqual;2\\ & 0& C_8\left(p_k\right)<2& \end{array}\right.,$ Wherein, I (p _kwith curve current point p in the cloudy discrete point sample graph of)=1 expression _kcorrespondence position is black pixel, C ₈(p _k) represent in cloudy discrete point sample graph with p _kblack number of pixels around correspondence position on 8 UNICOM directions.

3. a kind of form fit localization method based on discrete point sampling model according to claim 1, wherein, the matching rate operator utilizing positive discrete point sample graph to carry out described density Estimation is:

$\mathrm{\&gamma;}=\frac{N}{M}\&times;r^2$

In formula, M is summation of counting in region template, and N is region template discrete point summation in overlay area in discrete point sample graph, and r is the sample radius of negative and positive discrete point sampling model.

4. a kind of form fit localization method based on discrete point sampling model according to any one of claim 1-3, wherein, the structure conspicuousness that the described shape contour mould to extracting carries out preferably by calculating shape profile template matches feature realizes, wherein, the structure conspicuousness of described shape profile template matches feature is specifically calculated as follows:

If there is n class shape profile casting formwork target, and the effect of feature t is by arbitrary class ω _imake a distinction with other n-1 classes, then define this feature t for class ω _iarchitectural feature significance measure for:

$S_i^t=\frac{1}{n-1}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{S}_{\mathrm{ij}}^t,j\&NotEqual;i$

Wherein, and m _jfor class ω _iwith class ω _jthe average of corresponding templates matching characteristic t, with be its variance, a is proportional control factor, and b is index adjustment factor, and c is constant.

5. a kind of form fit localization method based on discrete point sampling model according to any one of claim 1-4, wherein, between shape to be extracted and described preferred shape contour mould, there is convergent-divergent, rotation geometry is out of shape, then need to carry out region conversion to shape contour mould, affine deformation Matrix Formula is:

$\left[\begin{array}{ccc}\mathrm{\&alpha;}& \mathrm{\&beta;}& (1-\mathrm{\&alpha;})\&times;\mathrm{center}.x-\mathrm{\&beta;}\&times;\mathrm{center}.y\\ -\mathrm{\&beta;}& \mathrm{\&alpha;}& \mathrm{\&beta;}\&times;\mathrm{center}.x+(1-\mathrm{\&alpha;})\&times;\mathrm{center}.y\end{array}\right]$

Wherein α=scale × cos (angle), β=scale × sin (angle), the scale coefficient of the corresponding affine deformation of scale, the coefficient of rotary of the corresponding affine deformation of angle, center.x and center.y represents horizontal ordinate and the ordinate of convergent-divergent or rotation center respectively.

6. a kind of form fit localization method based on discrete point sampling model according to any one of claim 1-5, wherein, under can using estimating target deformation condition, directly can determine deformation formula and then directly obtain corresponding contour mould, that is: first according to the priori of shape to be extracted, select from the center line extracted most possibly belong to shape to be extracted one or several center line as key central line, the size possible to shape contour mould, direction, position is estimated, then in cloudy discrete point sample graph corresponding key central line discrete point banded zone in coupling checking this shape contour mould estimated whether appropriate.

7., based on a form fit locating device for discrete point sampling model, it realizes images match by the shape contour mould extracted in discrete point sample graph, it is characterized in that, this device comprises:

First module, it for extracting shape contour mould from the cloudy discrete point sample graph in sample image;

Second module, it, for carrying out preferably to the shape contour mould extracted, obtains optimum shape contour mould: and

3rd module, it, for obtaining the cloudy discrete point sample graph of real image or positive discrete point sample graph, mates with corresponding optimum shape contour mould by it, thus recognition image;

Wherein, in described 3rd module, obtain with the matching rate operator of the density Estimation of corresponding optimum shape contour mould especially by the cloudy discrete point sample graph and/or positive discrete point sample graph that calculate described real image mating.

8. a kind of form fit locating device based on negative and positive discrete point sampling model according to claim 7, wherein, utilizes cloudy discrete point sample graph to carry out the matching rate operator of described density Estimation:

$\mathrm{\&lambda;}=\frac{N}{M}=\frac{\underset{k=1}{\overset{M}{\mathrm{\&Sigma;}}}S\left(p_k\right)}{M}$

In formula, M is summation of counting in shape contour mould, and N is the summation of counting that on shape contour mould, each point has response relation in discrete point sample graph banded zone, p _kbe current point on shape contour mould, arbitrary curve by the response relation of certain discrete point is $S\left(p_k\right)=\left\{\begin{array}{cccc}1& I\left(p_k\right)=1& \mathrm{or}& C_8\left(p_k\right)\&GreaterEqual;2\\ & 0& C_8\left(p_k\right)<2& \end{array}\right.,$ Wherein, I (p _kwith curve current point p in the cloudy discrete point sample graph of)=1 expression _kcorrespondence position is black pixel, C ₈(p _k) represent in cloudy discrete point sample graph with p _kblack number of pixels around correspondence position on 8 UNICOM directions.

9. a kind of form fit locating device based on negative and positive discrete point sampling model according to claim 7 or 8, wherein, the matching rate operator utilizing positive discrete point sample graph to carry out described density Estimation is:

$\mathrm{\&gamma;}=\frac{N}{M}\&times;r^2$

In formula, M is summation of counting in region template, and N is region template discrete point summation in overlay area in discrete point sample graph, and r is the sample radius of negative and positive discrete point sampling model.

10. a kind of form fit locating device based on negative and positive discrete point sampling model according to any one of claim 7-9, wherein, the structure conspicuousness that the described shape contour mould to extracting carries out preferably by calculating shape profile template matches feature realizes, wherein, the structure conspicuousness of described shape profile template matches feature is specifically calculated as follows:

If there is n class shape profile casting formwork target, and the effect of feature t is by arbitrary class ω _imake a distinction with other n-1 classes, then define this feature t for class ω _iarchitectural feature significance measure for:

$S_i^t=\frac{1}{n-1}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{S}_{\mathrm{ij}}^t,j\&NotEqual;i$

Wherein, and m _jfor class ω _iwith class ω _jthe average of corresponding templates matching characteristic t, with be its variance, a is proportional control factor, and b is index adjustment factor, and c is constant.