CN105913415A - Image sub-pixel edge extraction method having extensive adaptability - Google Patents

Abstract

The invention discloses an image sub-pixel edge extraction method having extensive adaptability, comprising steps of adopting an adaptive high-low threshold value calculation method, executing a local maximum central value choosing operation on a gradient image by combining with gradient direction information of a pixel point after obtaining a gradient image, establishing a relative coordinate with a random pixel point as an origin point, taking eight-neighborhood pixels around the origin point as local maximum central value choosing data samples, obtaining a comparison result between the adjacent neighborhoods according to a gradient direction, determining whether the current pixel position is a boundary point candidate position, adopting a Hessian matrix method based on a Steger surface fitting method to solve a sub-pixel position of the boundary point and connecting boundary points for forming a line to constitute a set of directed continuity points. Determination whether an extreme value of a local gradient magnitude is a boundary point can be made by combining with a specific threshold value; if the local extreme value is greater than a given threshold value, the local extreme value is marked as a boundary point, and if the local extreme value is smaller than a given threshold value, the local extreme value is a noise point or a background point The invention has great instantaneity.

Description

One has extensive adaptive image sub-pixel edge extracting method

Technical field

The present invention relates to image identification technical field, particularly relate to a kind of image sub-pixel edge extracting method.

Background technology

In machine vision, for carry out target location, measure, detect or Extraction of Geometrical Features etc. be required for right Target carries out the edge extracting of sub-pixel precision.In target positions, such as use the template matching of geometric properties Method needs to carry out template and target the edge extracting of sub-pixel precision；Need accurately to examine in measuring application The edge measuring object just can accurately measure；Detection application in, as optical character checking OCV, Edge defect detection etc. is required for stably detecting the sub-pixel edge of object.

Conventional Boundary extracting algorithm has Roberts operator, Sobel operator, Prewitt operator, Laplce Operator and Canny operator etc..The Boundary extracting algorithm of sub-pixel precision have space moments method, gray scale moments method, Zernike moments method and digital correlation etc..The Boundary extracting algorithm of other sub-pixel precision also includes that multinomial is intended Legal, ellipse fitting method, Gauss curved fitting process, Sigmoid curve-fitting method etc., Li Shuai etc. proposes one Kind of sub-pix detection algorithm based on Gauss curved matching, Sun Chengqiu etc. is at " the edge of a kind of sub-pixel precision Detection method " in propose to use Bezier edge model to carry out sub-pixel edge extraction, a dance is outstanding etc. proposes A kind of method of sub-pixel edge based on Sigmoid Function Fitting detection.(Chinese patent is public for patent documentation 1 The number of opening CN10465002A) disclose a kind of ellipse target sub-pixel edge extracted based on Sobel edge edge calmly Method for position, calculates elliptic geometry parameter by pixel edge, calculates sub-pixel edge by pixel edge.

Patent documentation 2 (China Patent Publication No. CN102737377A) discloses the sub-pix limit of a kind of improvement Edge extraction algorithm, first carries out the coarse positioning of pixel precision, utilizes edge image cutting target image to reduce lookup Scope, then extracts sub-pixel edge in the scope after reducing.Patent documentation 3 (China Patent Publication No. CN103530878A) disclose a kind of edge extracting method based on convergence strategy, use three kinds of traditional limits The result of edge extraction algorithm obtain reflection belong to edge may the ballot weight of degree, then analyze pixel with The maximum difference in luminance of neighborhood and the difference of minimum brightness difference, obtain the difference weight describing jump in brightness degree； Statistics goes center neighborhood variance distribution, obtains the marginal distribution weight of all pixels, carries out edge decision-making, Output edge image.Patent documentation 4 (China Patent Publication No. CN103886589A) disclose a kind of towards The automated high-precision edge extracting method of target, including model training stage and edge extracting stage.Patent Document 5 (China Patent Publication No. CN103955911A) discloses a kind of rim detection based on relative variation Method, including Image semantic classification and rim detection based on neural net method.Patent documentation 6 (Chinese patent Publication number CN104268857A) disclose a kind of fast sub-picture element rim detection and localization method, basic ideas It is first to obtain pixel edge position, then uses cosine look-up table to calculate sub-pixel edge point.Patent literary composition Offer 7 (China Patent Publication No. CN104268872A) and disclose a kind of based on conforming edge detection method. Patent documentation 8 (China Patent Publication No. CN104732536A) discloses a kind of based on improving morphologic Asia Pixel margin detection method, uses the morphological edge detector smoothed image marginal information improved, at thing Body edge contour utilizes Canny operator obtain the edge of Pixel-level, then fit to pixel edge produce The sub-pixel edge of product.Patent documentation 9 (China Patent Publication No. CN105005981A is public) has been opened a kind of The laser Rhizoma Dioscoreae (peeled) center extraction method of sub-pixel precision, by positioning in initial Rhizoma Dioscoreae (peeled) in the image after smooth The heart, then utilizes Gaussian function fitting to obtain Rhizoma Dioscoreae (peeled) width, the variance of recycling fitted Gaussian function and Gauss The parameters such as convolution kernel calculate Hessian matrix, go out according to Hessian matrix calculus in the sub-pix of laser Rhizoma Dioscoreae (peeled) Heart position.In the use of Hessian matrix method identical with the present invention, but in pixel accuracy positions calculating side There is the difference of essence in face, the suitability also causing two methods is entirely different.

But in industrial environment is applied, image is caused image degradation by the interference of all kinds of factors, bag Include very noisy, edge blurry etc., in low-quality image, how stably to detect the limit of sub-pixel precision Edge feature does not solves well.The Boundary extracting algorithm of traditional pixel precision is applied in industrial automation Required precision can not be met in the application such as 3C automated arm, electronic manufacture, industrial robot vision. The sub-pixel edge extraction algorithms such as space moments method, gray scale moments method, Zernike moments method and digital correlation are in detection All there is respective deficiency in precision, calculating speed and noise resisting ability aspect, is difficult in adaptation industrial environment tight Severe detection operating mode.

Patent documentation 1 can only extract the sub-pixel location of ellipse target, and versatility is not enough, and can not process mould Stick with paste the edge extracting problem of target.In method disclosed in patent documentation 3 distribution make use of Sobel, Canny and LoG operator carries out rim detection, then the result of three kinds of operator detections is weighted ballot statistics, according to The weight matrix of ballot obtains subpixel coordinates, and the method there is problems of that speed is slow, precision depends on power Weight matrix, very noisy, the edge extracting problem of the image such as fuzzy can not be solved.Patent documentation 4 and patent literary composition Method disclosed in 5 of offering uses the result of Canny and relative variation, uses the method for machine learning to carry out edge Extracting, its method speed is relatively slow, can not carry out stable edge extracting in low-quality image.Patent documentation 6 Disclosed method, on the basis of pixel coordinate coarse positioning, carries out sub-pixel edge inspection on 8 gradient directions Surveying, the method has and well calculates speed, but does not accounts for the process of very noisy and broad image.Patent It is the highest that document 7 and patent documentation 8 there is also computational efficiency, it is impossible to processes very noisy, the edge of broad image Extraction problem.Method disclosed in patent documentation 9 uses multi-level Gaussian convolution algorithm, and algorithm complex is high, right The zones of different non-linear shade change that the reasons such as illumination variation are formed cannot realize the light stripe centric line of robust and carry Taking, the method is only applicable to laser Rhizoma Dioscoreae (peeled) central line pick-up, it is impossible to realizes general picture edge characteristic and extracts.

Summary of the invention

It is an object of the invention to provide a kind of high speed, in high precision template matching of based on image edge information fixed Method for position, the method energy position of output template image sub-pixel precision in the target image, the anglec of rotation simultaneously Degree and the scaling factor, for target image occur displacement, rotate, scale, partial occlusion, optical illumination Dark change, uneven illumination is even, mixed and disorderly background etc. can realize location quick, stable, high-precision and identification. Present invention could apply to the occasion needing to be carried out target location and identification by machine vision: as robot draws Lead, the detection of semiconductor packages, electronic manufacture, Automated assembly, Product Visual, vision measurement, video with The fields such as track.

Method disclosed by the invention can stably detect in low-quality image that the edge of sub-pixel precision is special Property.

For reaching above-mentioned purpose, the present invention is achieved through the following technical solutions:

One has extensive adaptive image sub-pixel edge extracting method, comprises the following steps: step 1: Use the image blurring smothing filtering of variable dimension to Image semantic classification；Step 2: pretreated image is calculated First derivative, first ensures that the gradient magnitude obtained meets this error rate less than setting value α_p, image single order is led Number is obtained by target core convolved image space；The edge line of image at the crestal line of image first derivative, its In, crestal line is the set of adjacent continuous print local maximum in gradient image；Step 3: sieve at edge candidate point Edge extracting and the choosing principles of chain type threshold value it is applied to, it is achieved Pixel-level boundary position extracts during choosing, High-low threshold value uses two ways to obtain: external parameter input or adaptive threshold calculate；Step 4: After gradient image, for convenience of and be quickly found out the single pixel wide position of crestal line, in conjunction with the gradient side of pixel To information, gradient image is performed local maximum central value and select operation；Step 5: the extreme value of partial gradient amplitude Whether it is marginal point, needs to combine specific threshold and judge, be labeled as marginal point more than certain given threshold value, It is noise spot or background dot less than certain given threshold determination；Step 6: calculate the marginal position of sub-pixel precision； Step 7: marginal point is connected into curve, constitutes the set of one group of oriented continuity point.

In the most described step 2, if (x y) is performed edge extraction operation to image I After to obtain boundary point error rate be α_I, image size is n=w × h, then the probability of single-point detection mistake is α_p=1-(1-α_I)^{1 /n}, wherein α_IScope is between 0 to 1.0, and (x y) only has Gaussian noise and noise letter to image I Number variance is s_n；Utilize the substep characteristic of convolution, have a following equation:

$\left\{\begin{array}{c}{r}_x(x,y,\mathrm{\σ})={\left(I\right(x,y)*G(x,y,\mathrm{\σ}\left)\right)}^{\′}=(I*G{\left(x,\mathrm{\σ}\right)}^{\′})*G(y,\mathrm{\σ})\\ r_y(x,y,\mathrm{\σ})={\left(I\right(x,y)*G(x,y,\mathrm{\σ}\left)\right)}^{\′}=(I*G(x,\mathrm{\σ}\left)\right)*G{(y,\mathrm{\σ})}^{\′}\end{array}\right.,$

The gradient magnitude obtaining each point isMake the error rate of each point gradient magnitude less than setting Definite value α_p, i.e. meet equation: M (x, y, σ) >=c (σ), wherein,

$c=\frac{s_n}{2\sqrt{2\mathrm{\π}}{\mathrm{\σ}}^2}\sqrt{-2ln\[1-{(1-{\mathrm{\α}}_I)}^{1/n}\]},$

Variable in above formula is yardstick variable σ, and other variable is overall situation setup parameter.

As a further improvement on the present invention, described step 3 calculates high-low threshold value particularly as follows: first find Peak point (i, the H of histogram curve_i), non-zero accumulated value coordinate points is (j, H in last rectangular histogram_j),0≤i ＜ j≤255 and 0≤H_j＜ Hj ＜ 1.0, couples together above-mentioned 2, obtains straight line ax+by+c=0； Histogram curve coordinate points maximum distance position d to straight line is searched between i to j_max, the most satisfied d_{ma x}=arg max_k|ak+bH_k+ c |, this point coordinates (k, H_k) abscissa be first threshold value T_low=k；Connect From this point, to linear end point (j, H_j) reconnect α x+ β y+ λ=0 in alignment, in rectangular histogram On interval k to the j of curve, search curve maximum distance position D to straight line (α, β, λ)_max, the most satisfied D_max=arg max_t|α_t+βH_t+ λ |, this point coordinates (t, H_t) abscissa be labeled as second threshold value T_high=t.

As a further improvement on the present invention, described step 4 is particularly as follows: the gradient direction of any location of pixels For θ=tan^-1(f_y/f_x), the tangent tendency direction with crestal line；With any location of pixels as initial point, set up one Relative coordinate, taking eight neighborhood pixel around this point is that local maximum central value selects data sample, according to gradient Direction obtains the comparative result of neighborhood, determines whether current pixel position is boundary point position candidate.

As a further improvement on the present invention, described step 5 uses the dual threshold of Canny to set (T_low, T_high)； When local extremum G₀Higher than T_highTime, put p₀It it is marginal point；G₀Less than threshold value T_lowRepresent that current point is non-border Point attribute；Work as G₀Time between high-low threshold value, chain effect is had an effect, i.e. p₀The eight neighborhood of point is deposited At boundary point, then boundary point is confirmed as in current location.

As a further improvement on the present invention, described step 6 uses based on Steger curved surface fitting method Hessian matrix method asks for the sub-pixel location of marginal point, performs curved surface in the zonule of pixel edge point Interpolation algorithm f (r, c)=k of matching₀+k₁r+k₂c+k₃r²+k₄rc+k₅c²；Each the unknown to surface equation Number asks for single order and second dervative, is combined into Hessian matrix；Solve the eigenvalue of Hessian matrix and each From characteristic vector, wherein maximum characteristic vector corresponding to absolute feature value is the normal direction of marginal point (n_x,n_y)；Utilize the Taylor expansion of normal direction and surface equation, calculate the sub-pixel location of marginal point.

As a further improvement on the present invention, the border that should be noted that in the edge connection procedure of described step 7 The principle that connection is to be kept is to select recently and formed as far as possible straight line or the tendency of smooth curve, simultaneously Two curves of interconnection to be avoided the formation of, have for Wave curved and can only a song of existence anduniquess Line.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of the present invention；

Fig. 2 is at one dimensional image data and the schematic diagram of corresponding derivative results；

Fig. 3 is the schematic diagram that fixed size gaussian filtering coordinates Canny algorithm edge extracting result；

Fig. 4 is image-region matching and crestal line trend graph；

Fig. 5 is Threshold segmentation schematic diagram；

Fig. 6 (a) is central point pixel and eight neighborhood represents schematic diagram；

Fig. 6 (b) is eight neighborhood coordinate representation schematic diagram；

Fig. 7 is that currently some subsequent point is the direction of search and order schematic diagram；

Fig. 8 is that the image sub-pixel edge with very noisy extracts example schematic；

Fig. 9 is image mutative scale broad image testing result schematic diagram；

The testing result contrast schematic diagram that Figure 10 is the method for the present invention and commercial software obtains.

Specific embodiments

Combine accompanying drawing below by detailed description of the invention the present invention is described in further detail.

Rim detection is the algorithm being widely used in image procossing, and in machine vision technique, very multi-operator is all Need to know based on good edge extracting result, such as geometric templates coupling, straight-line detection, loop truss, character Not, defects detection, dimensional measurement etc..The invention provides a kind of energy stable detection very noisy image or fuzzy The method of the strong image border of dimensional variation, the method can provide the marginal position of sub-pixel precision, marginal point Annexation, the length information of marginal point.Rim detection efficiency is extremely efficient, is very suitable for regarding at machine Feel real-time system is applied.The present invention can provide important foundation for the location in machine vision, technology of measuring.

The method flow diagram of the present invention as shown in Figure 1, comprises the following steps: step 1: Image semantic classification, To image smoothing filtering technique；Step 2: nanoscale regime single order discrete kernel convolved image；Step 3: self adaptation height Threshold calculations；Step 4: approximate gradient direction calculating local maximum central value select；Step 5: pixel limit Boundary's point judges to select；Step 6: calculate the marginal position of sub-pixel precision；Step 7: same alike result boundary point Order connects.

Below each step is specifically described.

1. Image semantic classification

Before searching marginal point, an edge model meeting specified conditions need to be set up.Most edges Detection algorithm, such as Marr, Hildreth, Poggio, Canny etc., the marginal position of definition is at gradation of image The position of sudden change, i.e. first derivative amplitude data are same equal to zero higher than the position of certain threshold value or second dervative Time be not smooth flex point (flat inflection point), meet condition g ' (x, y) g^m(x, y) ＜ 0.In accompanying drawing 2, g (s) is One-dimensional intensity profile, in figure, g ' (s) is the first derivative curve of one-dimensional intensity profile, and " (s) is one-dimensional intensity profile to g Second derivative curves.Single order, second dervative can represent picture edge characteristic, but first derivative has meter Calculate the advantage that speed is fast, anti-noise ability is strong, the present invention uses first derivative as the foundation judging edge.

Curve g (s) of accompanying drawing 2 display represents initial data, this figure can assume that step edge model is Ku (x)+h, wherein k is unknown gradient magnitude, and h represents the gray values of background image, and u (x) is intensity profile Curvilinear equation.Edge extracting uses the purpose of pretreatment be all possible marginal position is all processed after Close to edge model g (s) of accompanying drawing 2, it also it is step process requirement 1. in method flow.Image semantic classification leads to Cross Gauss to be filtered into, and the object processed include: uncertain types noise interferogram picture, different reason and The broad image of degree, non-step edge model.The Gaussian Blur core of two dimensional image is defined as follows:

$G(x,y,\mathrm{\σ})=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-(x^2+y^2)/2{\mathrm{\σ}}^2}$

Unknown variable dimension information therein is curve variances sigma parameter.For being adapted to different types of image, it is ensured that can To solve the problem run in accompanying drawing 3, variable dimension is image blurring, and smothing filtering is used to do the pre-place of image Reason means.Need solution is i.e. when meeting subregion marginal point performance, and additionally part edge point is tested Measure non-single pixel wide boundary point.

2. single order yardstick gaussian kernel derivation Image edge gradient

After image pretreatment operation, it is present in the definition at gradation of image value mutation according to marginal point, simultaneously Edge line is also crestal line (ridge) place in image first derivative, as shown in Figure 4.In crestal line is gradient image The set of the local maximum of adjacent continuous, is also the place of boundary curve simultaneously.

Image when generating, due to a variety of causes can cause obscurity boundary (as stationary lens to the different refractions of light, Non-parallel light forms shade in boundary, edge self transition obscures) or introduce noise signal (such as Gauss Noise), the present invention can overcome the problems referred to above to obtain single pixel wide marginal point.(x y) is performed edge to set image I Obtaining boundary point error rate after extracting operation is α_I, image size is n=w × h, then the probability of single-point detection mistake For α_p=1-(1-α_I)^{1/ n}；Wherein α_IScope is between 0 to 1.0.When calculating image first derivative, first Guarantee that the gradient magnitude obtained meets this error rate less than setting value α_p.Image first derivative is rolled up by target core Long-pending image space obtains, and utilizes the substep characteristic of convolution, has a following equation:

$\left\{\begin{array}{c}{r}_x(x,y,\mathrm{\σ})={\left(I\right(x,y)*G(x,y,\mathrm{\σ}\left)\right)}^{\′}=(I*G{\left(x,\mathrm{\σ}\right)}^{\′})*G(y,\mathrm{\σ})\\ r_y(x,y,\mathrm{\σ})={\left(I\right(x,y)*G(x,y,\mathrm{\σ}\left)\right)}^{\′}=(I*G(x,\mathrm{\σ}\left)\right)*G{(y,\mathrm{\σ})}^{\′}\end{array}\right.---\left(1\right)$

The gradient magnitude of each point is

If (x, y) only have Gaussian noise and noise variance signal is s to image I_n, function U represents the positive district of Gaussian function Between semi-function, its partial derivative expression formula is:

$p_U\left(u\right)=\frac{2}{\sqrt{2\mathrm{\π}}s}{e}^{\frac{-u^2}{2s^2}},u\∈\[0,\mathrm{\∞})---\left(2\right)$

Wherein the relation of picture signal variance and filtering signal variance is s=| | G ' (x, y, σ) | |₂s_n, function f is differential Homeomorphic Maps (diffeomorphism), and have V=f (U), then the partial derivative of function V is:

$p_V\left(v\right)p_U\left(f^{-1}\right(v\left)\right)\left|\frac{d}{dv}{f}^{-1}\left(v\right)\right|,v\∈f\left(A\right)---\left(3\right)$

Constructed fuction f (u)=u², associating (2) formula and (3) formula, obtain following formula:

$p_V\left(v\right)=\frac{1}{\sqrt{\mathrm{\π}v}s}{e}^{-v/2s^2},v\∈\[0,\mathrm{\∞})---\left(4\right)$

The gradient of each axle in convolution (1), and substitute into function (4), has a following expression:

$p_{V_1+V_2}\left(v\right)=p\[r_x^2+r_y^2=v\]={\∫}_0^v{p}_V\left(v^{\′}\right)p_V(v-v^{\′}){\mathrm{dv}}^{\′}---\left(5\right)$

Solve (5) formula, obtainV ∈ [0, ∞).The marginal position of guarantee each point makes a mistake Probability is less than α_p, probability density function (5) formula is done Integral Processing and obtains probit.Setting key parameter becomes Amount for c, has a following expression:

${\∫}_c^{\mathrm{\∞}}{p}_{V_1+V_2}\left(v\right)dv={\mathrm{\α}}_p---\left(6\right)$

The L of Gaussian function first derivative₂Distance is:In conjunction with above-mentioned associated expression, The expression formula of parameter c is:

$c=\frac{s_n}{2\sqrt{2\mathrm{\π}}{\mathrm{\σ}}^2}\sqrt{-2ln\[1-{(1-{\mathrm{\α}}_I)}^{1/n}\]}---\left(7\right)$

The variable of its Chinese style (7) is σ, is i.e. yardstick variable, and other variable is overall situation setup parameter.By (5) (6) two formula Can obtain, variable metric algorithm it is critical only that the error rate so that each point gradient magnitude is less than setting value α_p, the most satisfied Equation: M (x, y, σ) >=c (σ).

3. self adaptation high-low threshold value calculates

The present invention is applied to edge extracting and the choosing principles of chain type threshold value in edge candidate point screening process, Realizing Pixel-level boundary position to extract, i.e. in flow process, step is 3..The present invention uses two ways to set height threshold Value: external parameter input and adaptive threshold calculate.The histogram curve of image gradient information synthesis exists one Individual obvious peak point, and histogrammic numerical value drastically reduces until dropping to null value after peak point.

The invention discloses a kind of simple and quick lookup high-low threshold value method.First histogram curve is found Peak point (i, H_i), it is additionally that in last rectangular histogram, non-zero accumulated value coordinate points is (j, H_j) (0≤i ＜ j≤255) And 0≤H_j＜ H_i＜ 1.0), above-mentioned 2 are coupled together, obtains straight line ax+by+c=0.At i to j Between search histogram curve coordinate points to the maximum distance position of straight line, the most satisfied d_max=arg max_k|ak+bH_k+c|.This point coordinates (k, H_k) abscissa be first threshold value T_low=k.Connect From this point, to linear end point (J, H_j) reconnect α x+ β y+ λ=0 in alignment, in rectangular histogram On interval k to the j of curve, search the curve maximum distance position to straight line (α, β, λ), the most satisfied D_max=arg max_t|αt+βH_t+λ|.This point coordinates (t, H_t) abscissa be labeled as second threshold value T_high=t. The simple performance of calculation is as shown in Figure 5.

4. approximate gradient direction calculating and local maximum central value select

After obtaining gradient image, for convenience of and be quickly found out the single pixel wide position of crestal line, in conjunction with pixel Gradient direction information to gradient image perform local maximum central value select operation, be flow chart step 4.. Arbitrarily the gradient direction of location of pixels is θ=tan^-1(f_y/f_x), the tangent tendency direction with crestal line.With any pixel Position is initial point, sets up a relative coordinate, and taking eight neighborhood pixel around this point is the choosing of local maximum central value Select data sample, obtain the comparative result of neighborhood according to gradient direction, determine whether current pixel position is limit Boundary's point position candidate.

During edge extracting, non-maxima suppression is the effective means of fast selecting local maximum.First Before this gradient direction (0 °～180 °) with 22.5 ° as step-length, was divided into some regions, such as Fig. 6 (a) institute Show.Wherein both angled complimentary of A and A ' are considered as same group.The invention provides a kind of Fast Field direction Algorithm for estimating, judge the most easily currently to put gradient direction in eight neighborhood direction towards.If it is current Point p₀Gradient direction θ beWherein variable is respective gradient derivative numerical value, and angular range sets Be scheduled between 0 ° to 90 °, the most only consider parameter be on the occasion of state.When θ is less than 22.5 °, n_y＜ n_xTan (22.5 °), Gradient direction is seated in the A scope of Fig. 6 (a), then p₀Eight neighborhood direction be the G in Fig. 6 (b)_R；Work as θ During more than 67.5 °, n_y＞ n_xTan (67.5 °), gradient direction is seated in the C scope of Fig. 6 (a), then p₀Eight Neighborhood direction is the G in Fig. 6 (b)_T；When θ scope is between 22.5 ° and 67.5 °, n_y≥n_xTan (22.5 °) and n_y≤n_xTan (67.5 °), gradient direction is seated in the D scope of Fig. 6 (a), then p₀Eight neighborhood direction be Fig. 6 G in (b)_TR.If p₀The gradient direction G of point₀Fall in the range of the A of Fig. 6 (a), then by eight neighborhood side G in_RWith G_LTwo complementary direction are respectively labeled as G₊And G_-.The criterion of local maximum is: G₀＞ G₊And G₀≥G_-Or G₀≥G₊And G₀＞ G_-, the most currently put p₀It it is a partial gradient amplitude maximum Value position.In multilevel iudge, if two to compare symbol be all " > " number, then at equal gradient magnitude, meeting The situation not having extreme value occurs；If be all " >=" number time, the position that the most all amplitudes are equal all can be confirmed to be pole Value.

5. pixel boundary point judges to select

Whether the extreme value of partial gradient amplitude is marginal point, needs to combine specific threshold and judges, gives more than certain That determines threshold value is labeled as marginal point, is noise spot or background dot, i.e. flowchart steps less than certain given threshold determination The most 5..The dual threshold of Canny sets (T_low, T_high) be used in the present invention.When local extremum G₀Higher than T_high Time, put p₀It it is marginal point；G₀Less than threshold value T_lowRepresent that current point is non-boundary point attribute；Work as G₀Between height Time between threshold value, chain effect is had an effect, i.e. p₀There is boundary point in the eight neighborhood of point, then current location Confirm as boundary point.

6. calculate the marginal position of sub-pixel precision

In being normally applied, the boundary point positional precision of Pixel-level can meet demand, but needs in some applications Wanting higher edge definition position, i.e. sub-pixel edge position, the step of flow chart is 6..Use based on Steger The Hessian matrix method of curved surface fitting method asks for the sub-pixel location of marginal point, little at pixel edge point Interpolation algorithm f (r, c)=k of surface fitting is performed in region₀+k₁r+k₂c+k₃r²+k₄rc+k₅c²；To song Each unknown number of face equation asks for single order and second dervative, is combined into Hessian matrix；Solve Hessian square The eigenvalue of battle array and respective characteristic vector, wherein maximum characteristic vector corresponding to absolute feature value is limit Normal direction (the n of edge point_x,n_y)；Utilize the Taylor expansion of normal direction and surface equation, calculate the Asia of marginal point Location of pixels.The matrix expression of surface equation coefficient is as follows.

$\begin{array}{ccc}{k}_1=\frac{1}{6}\left\{\begin{array}{ccc}-1& -1& -1\\ 0& 0& 0\\ 1& 1& 1\end{array}\right\}& k_2=\frac{1}{6}\left\{\begin{array}{ccc}-1& 0& 1\\ -1& 0& 1\\ -1& 0& 1\end{array}\right\}& k_3=\frac{1}{6}\left\{\begin{array}{ccc}1& 1& 1\\ -2& -2& -2\\ 1& 1& 1\end{array}\right\}\end{array}$

$\begin{array}{cc}{k}_4=\frac{1}{6}\left\{\begin{array}{ccc}1& 0& -1\\ 0& 0& 0\\ -1& 0& 1\end{array}\right\}& k_5=\frac{1}{6}\left\{\begin{array}{ccc}1& -2& 1\\ 1& -2& 1\\ 1& -2& 1\end{array}\right\}\end{array}$

7. same alike result boundary point is linked in sequence

Until now, the marginal information detected is discrete, unordered, isolated point, but a lot of later stage Application is it is desirable that there is successional boundary point set (curve), and flow chart step is 7..Marginal point is connected into song Line, constitutes the set of one group of oriented continuity point.Edge connection procedure it should be noted that, contour connection to be protected The principle held is to select recently and form straight line or the tendency of smooth curve as far as possible.The most also to avoid Form two curves interconnected, Wave curved is had and can only a curve of existence anduniquess.

Edge connect the edge point position that available condition is image space and this eight neighborhood boundary point exist with No.Curve starting point starts search from the upper left corner, detects that first marginal point is i.e. defined as starting position.Should The order in point search direction is preferentially to search positive direction (i.e. such as the G of accompanying drawing 7_R, G_B, G_L, G_TWhether have on) Meet the boundary point of condition, otherwise search folk prescription to (other direction in accompanying drawing 7).Similar direction (positive direction, Folk prescription to) in, candidate point select according to order counterclockwise.If P₀It is current point, { P_i}_{I=R, B, T, BR, RT} It is candidate point, the subpixel coordinates of each pointAnd gradient directionFor known conditions.Given One evaluation functionThe neighborhood point selecting score value minimum is considered as next phase Adjoint point.Circulation change also replaces current point, until the edge run on non-edge point or other boundary curve Point just terminates the search of current curves.After current direction has been searched for, then start opposite direction from the origin of curve and look into Look for, until terminal.

In order to verify the effectiveness of the open method of the present invention, distribution uses very noisy (seeing accompanying drawing 8 left part) Edge extracting test, accompanying drawing 8 and the pars intermedia of accompanying drawing 9 is carried out with broad image (seeing accompanying drawing 9 left part) Being divided into the result of traditional rim detection, right part is testing result of the present invention, it can be seen that the present invention is public The method opened can stably detect edge feature in very noisy and broad image.Accompanying drawing 10 is the present invention's The edge detection results that method obtains is right with the Germany edge detection results that obtains of commercialization machine vision software Ratio effect, in figure "+" it is the edge result that detects of the present invention, " ο " is external business software testing result, As can be seen from the figure the method in the present invention can detect the marginal information of image more than enoughly.

The present invention proposes a kind of method stably extracting image sub-pixel edge feature in adverse circumstances, Use self adaptation high-low threshold value computational methods, after obtaining gradient image, for convenience of and be quickly found out crestal line Single pixel wide position, gradient image execution local maximum central value is selected by the Gradient direction information in conjunction with pixel Selecting operation, with any location of pixels as initial point, set up relative coordinate, taking eight neighborhood pixel around this point is office The very big central value in portion selects data sample, obtains the comparative result of neighborhood according to gradient direction, determines current picture Whether element position is boundary point position candidate.Whether the extreme value of partial gradient amplitude is marginal point, needs to combine Specific threshold judges, is labeled as marginal point, less than certain given threshold determination for making an uproar more than certain given threshold value Sound point or background dot.Hessian matrix method based on Steger curved surface fitting method is used to ask for the Asia of marginal point Location of pixels.Last just marginal point connects into curve, constitutes the set of one group of oriented continuity point.Achieve The orderly edge feature information of sub-pixel precision is extracted in very noisy and broad image.The inventive method has Fabulous real-time, it is possible to be applied in the application in real time of Vision Builder for Automated Inspection.

Above content is to combine concrete preferred implementation further description made for the present invention, no Can assert the present invention be embodied as be confined to these explanations.Common for the technical field of the invention For technical staff, without departing from the inventive concept of the premise, it is also possible to make some simple deductions or replace Change, all should be considered as belonging to protection scope of the present invention.

Claims (7)

1. one kind has extensive adaptive image sub-pixel edge extracting method, it is characterised in that: described method bag Include following steps:

Step 1: use the image blurring smothing filtering of variable dimension to Image semantic classification；

Step 2: pretreated image is calculated first derivative, first ensures that the gradient magnitude obtained meets and is somebody's turn to do Point error rate is less than setting value α_p, image first derivative is obtained by target core convolved image space；Image Edge line at the crestal line of image first derivative, wherein, crestal line is adjacent continuous print office in gradient image The set of portion's maximum；

Step 3: be applied to edge extracting and the choosing principles of chain type threshold value in edge candidate point screening process, Realize Pixel-level boundary position extract, high-low threshold value use two ways obtain: external parameter input or Adaptive threshold calculates；

Step 4: after obtaining gradient image, for convenience of and be quickly found out the single pixel wide position of crestal line, in conjunction with The Gradient direction information of pixel performs local maximum central value to gradient image and selects operation；

Step 5: whether the extreme value of partial gradient amplitude is marginal point, needs to combine specific threshold and judges, greatly It is labeled as marginal point in certain given threshold value, is noise spot or background dot less than certain given threshold determination；

Step 6: calculate the marginal position of sub-pixel precision；

Step 7: marginal point is connected into curve, constitutes the set of one group of oriented continuity point.

Method the most according to claim 1, it is characterised in that: in described step 2, if (x y) is held image I Obtaining boundary point error rate after row edge extraction operation is α_I, image size is n=w × h, then single-point detection The probability of mistake is α_p=1-(1-α_I)^1/n, wherein α_IScope is between 0 to 1.0, and (x y) only has height to image I This noise and noise variance signal are s_n；Utilize the substep characteristic of convolution, have a following equation:

$\left\{\begin{array}{c}{r}_x(x,y,\mathrm{\σ})={\left(I\right(x,y)*G(x,y,\mathrm{\σ}\left)\right)}^{\′}=(I*G{\left(x,\mathrm{\σ}\right)}^{\′})*G(y,\mathrm{\σ})\\ r_y(x,y,\mathrm{\σ})={\left(I\right(x,y)*G(x,y,\mathrm{\σ}\left)\right)}^{\′}=(I*G(x,\mathrm{\σ}\left)\right)*G{(y,\mathrm{\σ})}^{\′}\end{array}\right.,$

The gradient magnitude obtaining each point isThe error rate making each point gradient magnitude is less than Setting value α_p, i.e. meet equation: M (x, y, σ) >=c (σ), wherein,

$c=\frac{s_n}{2\sqrt{2\mathrm{\π}}{\mathrm{\σ}}^2}\sqrt{-2ln\[1-{(1-{\mathrm{\α}}_I)}^{1/n}\]},$

Variable in above formula is yardstick variable σ, and other variable is overall situation setup parameter.

Method the most according to claim 1, it is characterised in that: described step 3 calculates high-low threshold value concrete For: first find peak point (i, the H of histogram curve_i), non-zero accumulated value coordinate in last rectangular histogram Point is (j, H_j), 0≤i ＜ j≤255 and 0≤H_j＜ Hj ＜ 1.0, couples together above-mentioned 2, obtains one Straight line ax+by+c=0；The histogram curve coordinate points maximum distance position to straight line is searched between i to j d_max, i.e. meet d_max=argmax_k|ak+bH_k+ c |, this point coordinates (k, H_k) abscissa be first Threshold value T_low=k；The most from this point, to linear end point (j, H_j) reconnect in alignment α x+ β y+ λ=0, on interval k to the j of histogram curve, searches the curve maximum to straight line (α, β, λ) Distance and position D_max, meet D equally_max=argmax_t|αt+βH_t+ λ |, this point coordinates (t, H_t) horizontal seat It is labeled as second threshold value T_high=t.

Method the most according to claim 1, it is characterised in that: described step 4 is particularly as follows: any pixel position The gradient direction put is θ=tan^-1(f_y/f_x), the tangent tendency direction with crestal line；With any location of pixels it is Initial point, sets up a relative coordinate, and taking eight neighborhood pixel around this point is that local maximum central value selects number According to sample, obtain the comparative result of neighborhood according to gradient direction, determine whether current pixel position is border Point position candidate.

Method the most according to claim 4, it is characterised in that: described step 5 uses the dual threshold of Canny Set (T_low, T_high)；When local extremum G₀Higher than T_highTime, put p₀It it is marginal point；G₀Less than threshold value T_lowTable Show that current point is non-boundary point attribute；Work as G₀Time between high-low threshold value, chain effect is had an effect, I.e. p₀There is boundary point in the eight neighborhood of point, then boundary point is confirmed as in current location.

Method the most according to claim 1, it is characterised in that: described step 6 uses based on Steger curved surface The Hessian matrix method of approximating method asks for the sub-pixel location of marginal point, little at pixel edge point Interpolation algorithm f (r, c)=k of surface fitting is performed in region₀+k₁r+k₂c+k₃r²+k₄rc+k₅c²；Right Each unknown number of surface equation asks for single order and second dervative, is combined into Hessian matrix；Solve Hessian The eigenvalue of matrix and respective characteristic vector, wherein maximum characteristic vector corresponding to absolute feature value is i.e. Normal direction (n for marginal point_x,n_y)；Utilize the Taylor expansion of normal direction and surface equation, calculate limit The sub-pixel location of edge point.

Method the most according to claim 1, it is characterised in that: the edge connection procedure of described step 7 needs It is noted that a contour connection principle to be kept be to select recently and form straight line or smooth as far as possible The tendency of curve, two curves of interconnection to be avoided the formation of, Wave curved is had and only One curve of energy existence anduniquess.