CN105574864A - Angle accumulation-based self-adapted corner point detection method - Google Patents

Abstract

The invention discloses an angle accumulation-based self-adapted corner detection point method. The method comprises the following steps: after obtaining the image edge, defining the concept of an edge point reflecting the local features, calculating the angle accumulation value of the edge point, and taking the angle accumulation value as corner point initial response; constructing the local self-adapted threshold values of candidate corner points to remove the round corner points; and obtaining the feature values of the corner points, and constructing the global threshold values to remove the false corner points so as to obtain the final corner point detection result.

Description

Based on the self-adaptive angular-point detection method that angle is cumulative

Technical field

The present invention relates to a kind of self-adaptive angular-point detection method cumulative based on angle, can be applicable to the aspects such as target identification, Stereo matching, motion tracking, 3D reconstruction, belong to technical field of image processing.

Background technology

Angle point is the important local feature of one of image, and angle point effectively reduces the data volume of information while retention body weight wants characteristic information, and the operand of image procossing is greatly reduced.In addition angle point has concentrated the much important shape information on image, and has rotational invariance, and therefore angle point is hardly by the impact of illumination condition.In the image registration, the field such as image understanding and pattern-recognition of feature based, Corner Detection tool is of great significance.

Up to the present, have and be suggested based on the angular-point detection method at edge in a large number, these methods are mainly divided into following three classes:

(1) based on the Corner Detection of CSS

Conventional method based on the Corner Detection of CSS is: first under larger yardstick, calculate marginal point curvature, getting Local modulus maxima is candidate angular, removes pseudo-angle point, finally accurately locate angle point under low yardstick by global threshold.These class methods are faced with following deficiency: one is the character due to curvature self, need to calculate single order second derivative, make the change of algorithm edge local very responsive; Two is select suitable Gauss's yardstick still very difficult.So the accuracy of such angular-point detection method is lower.

(2) based on the Corner Detection of CPDA

Angular-point detection method based on CPDA adopts point to add up to chordal distance and is used as the feature interpretation of marginal point, and the calculating due to this feature is based on Euclidean distance completely, so algorithm avoids CSS algorithm Problems existing.But these class methods are faced with following problem: because the eigenwert of dissimilar angle point differs greatly, Corner Feature value difference of the same type under different chord lengths is also very large, thus this type of algorithm must be normalized Corner Feature value (maximal value divided by same curve), but this can bring new problem: as curve existed sharp-pointed angle point, its eigenwert is very large, then the angle point that eigenwert is less effectively can not distinguish with pseudo-angle point after normalization; And if for example curve does not exist true angle point, then some pseudo-angle points will be very large by the eigenwert after normalization, can be erroneously detected as angle point.

(3) based on the Corner Detection Algorithm of Angle

Based on the Corner Detection of Angle, the angle adopting the end points line of marginal point and a string to form is used as feature interpretation, the size of Angle not only can reflecting edge point patterns, and the permanent set of Angle value is within certain scope: [0, π], so do not need to be normalized, avoid based on the shortcoming existing for the method for CPDA.But these class methods remain a lot of not enough: can not detect contiguous angle point, effectively cannot remove circular angle point, false angle point.

The invention belongs to the angular-point detection method of (3) class based on Angle, be called based on the cumulative self-adaptive angular-point detection method of angle.

Summary of the invention

In order to overcome the existing defect based on existing in the angular-point detection method at edge, the problems such as too high in angle point error rate, Corner character precision is low, the present invention proposes a kind of self-adaptive angular-point detection method cumulative based on angle, to orient angle point fast and accurately.

For achieving the above object, the present invention proposes based on the cumulative self-adaptive angular-point detection method of angle, institute's extracting method, behind acquisition image border, defines the concept that the angle of reflecting edge point local feature is cumulative, and edge calculation point angle accumulated value, in this, as angle point initial response; Then the local auto-adaptive threshold value constructing candidate angular removes circular angle point; Then obtain Corner Feature value, and construct global threshold and remove false angle point, obtain final Corner Detection result.

The self-adaptive angular-point detection method concrete steps that the present invention is based on angle cumulative comprise following step:

(1) adopt canny operator to extract edge, carry out smooth curve by a wicket gaussian kernel;

(2) adopt L value be 6 string carry out the angle accumulated value of edge calculation point, minimalization point is candidate angular;

In step (2), first define the concept that angle is cumulative: the pixel on edge is denoted as P ₁, P ₂..., P _k, P _k+1..., P _n, n is curve up contour point number, for marginal point P each on curve _k, string C _lfrom P _k-L+1p _k+1position starts, and connects string P _k-L+1p _kwith string P _kp _k+1, calculate the angle theta of these two strings _{k, k-L+1}, then string C _lmove to right a pixel to P _{k-L+ 2}p _k+2, calculate string P _k-l+2p _kwith string P _kp _k+2angle theta _{k, k-L+2}, by parity of reasoning, until string C _lmove to P _k-1p _k+L-1till position, calculate all angle sums, be marginal point P _kthe angle accumulated value θ at place _l(k).

${\mathrm{\θ}}_L\left(k\right)=\underset{i=k-L+1}{\overset{k-1}{\Σ}}{\mathrm{\θ}}_{k,i}---\left(1\right)$

Then a shorter string C is got _l(L=6), the θ of each pixel on calculated curve _lk () value, due to θ _{k, k-L+j}scope be [0, π], so θ _lk the scope of () is [0, (L-1) π], in order to make experimental data more directly perceived, angle accumulated value being carried out formula (2) conversion and obtaining θ ' _lk (), its scope is [0,1], and minimalization point is incorporated to candidate angular set.

${\mathrm{\θ}}_L^{\′}\left(k\right)=\frac{{\mathrm{\θ}}_L\left(k\right)}{(L-I)\mathrm{\π}}---\left(2\right)$

(3) the local auto-adaptive angle threshold constructing candidate angular removes circular angle point;

In step (3), by the marginal point θ ' in angle point neighborhood _lk () value constructs adaptive threshold T _k, shown in (3), if the θ ' of candidate angular _lwhat k () was greater than adaptive threshold is circular angle point, rejects from candidate angular set.

$T_k=x\×\frac{1}{L_1+L_2+1}\underset{i=k-L_1}{\overset{k+L_2}{\mathrm{\Σ}}}{{\mathrm{\θ}}^{,}}_L\left(i\right)---\left(3\right)$

Wherein, x is scale-up factor, and value is 0.96.L ₁and L ₂be respectively left and right Size of Neighborhood, L ₁=min (L, left), L ₂=min (L, right), L=6, left, right are respectively a P _kcount in edge to left and right limit maximum point.

(4) string adopting L value to be respectively 10,20,30 to candidate angular carrys out the eigenwert of calculated candidate angle point, and constructs global threshold and remove pseudo-angle point, obtains final angle point set.

In step (4), length is adopted to be respectively the string of 10,20,30, such as formula (4), the eigenwert Φ (k) that product obtains candidate angular is carried out to three, then global threshold T is constructed by formula (5), wherein angle is the obtuse angle value of artificial setting, if the eigenwert Φ (k) of candidate angular is less than global threshold, be then false angle point, reject from candidate angular set.

$\mathrm{\Φ}\left(k\right)=\underset{j=1}{\overset{3}{\Π}}(1-{\mathrm{\θ}}_{L_j}^{\′}(k\left)\right)---\left(4\right)$

$T={(1-\frac{angle}{180})}^3---\left(5\right)$

Compared with the conventional method, advantageous of the present invention exists:

(1) only only used image edge structure information, method is short and sweet;

(2) to the curve of each in image, cumulative as its eigenwert using each point angle on edge, make Corner Detection robustness stronger;

(3) construct local auto-adaptive threshold value, can effectively remove circular angle point;

(4) obtained the eigenwert of angle point by ' many strings ' method, true angle point and false angle point are more effectively made a distinction;

(5) there is the ability of stronger opposing noise.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the inventive method;

Fig. 2 is the original image carrying out Corner Detection;

Fig. 3 carries out the design sketch after edge extracting to original image;

Fig. 4 calculates the cumulative principle schematic of angle;

Fig. 5 is the candidate angular extracted in original image;

Fig. 6 is the principle schematic removing circular angle point;

Fig. 7 is the lab diagram of selection percentage coefficient;

Fig. 8 is lab diagram original image being removed to false angle point;

Fig. 9 is the final angle point figure detected in original image.

Embodiment

Technical scheme for a better understanding of the present invention, below in conjunction with drawings and Examples, the present invention is described in further detail.

The inventive method carries out the process flow diagram of Corner Detection as shown in Figure 1 to image.

Rim detection is carried out, Gaussian smoothing after reading in image; The angle accumulated value of the last point of edge calculation; Structure local auto-adaptive threshold value removes circular angle point; Calculate the eigenwert of angle point, structure global threshold removes false angle point.

Embodiments of the invention are undertaken by step shown in accompanying drawing 1, specific as follows:

1, reading images, rim detection, Gaussian smoothing

As shown in Figure 2, process image with edge detection operator, carry out Gaussian smoothing after obtaining binaryzation edge, Fig. 3 is to the design sketch after Gaussian smoothing to the original image read.

2, candidate angular initial sets

Fig. 4 is the principle schematic of edge calculation point angle accumulated value, contrasts Fig. 4 below and is specifically described angle point is cumulative.

In Fig. 4, on the local configuration of image border, P ₁, P ₂..., P _nn marginal point on curve.Edge calculation point P _kthe cumulative principle of angle be: by a long string C for L _lfrom P _k-L+1p _k+1position starts, and connects string P _k-L+1p _kwith string P _kp _k+1, calculate the angle theta of these two strings _{k, k-L+1}, shown in (1), then string C _lmove to right a pixel to P _k-L+2p _k+2, calculate string P _{k-l+ 2}p _kwith string P _kp _k+2angle theta _{k, k-L+2}, by parity of reasoning, until string C _lmove to P _k-1p _k+L-1till position, calculate all angle sums, be marginal point P _kthe angle accumulated value θ at place _lk (), shown in (2).

${\mathrm{\θ}}_{k,k-L+j}={\cos }^{-1}\frac{{\left|P_{k-L+j}{P}_k\right|}^2+{\left|P_k{P}_{k+j}\right|}^2-{\left|P_{k-L+j}{P}_{k+j}\right|}^2}{2*\left|P_{k-L+j}{P}_k\right|*\left|P_k{P}_{k+j}\right|}---\left(1\right)$

${\mathrm{\θ}}_L\left(k\right)=\underset{i=k-L+1}{\overset{k-1}{\Σ}}{\mathrm{\θ}}_{k,i}---\left(2\right)$

The foundation of institute's extracting method is: at corner point, and the support angle that two, left and right string is formed is smaller, and can be partially formed a minimal value.Herein by mobile C _lobtain more strut angle, obtain angle accumulated value by summation, robustness is higher, can better reflect marginal point feature.Due to θ _{k, k-L+j}scope be [0, π], so θ _lk the scope of () is [0, (L-1) π], in order to make experimental data more directly perceived, angle accumulated value being carried out formula (3) conversion and obtaining θ ' _lk (), makes its scope become [0,1]:

${\mathrm{\θ}}_L^{\′}\left(k\right)=\frac{{\mathrm{\θ}}_L\left(k\right)}{(L-I)\mathrm{\π}}---\left(3\right)$

Select local minizing point be candidate angular as shown in Figure 5, Fig. 5 is the angle point design sketch detected from original image, with ' ' indicate be candidate angular.

3, circular angle point is removed

Circular object can't produce angle point clearly in human eye, so in the identification of image, circular angle point is so not meaningful.But the θ ' of circular angle point _lk () value is similar to some true angle points, and also can form local minimum, so carry out in the extraction of candidate angular in 2 steps, circular angle point is also selected into interior.In order to reject circular angle point, next algorithm construction local auto-adaptive threshold value of carrying removes circular angle point.

Find through experiment, as shown in Figure 6, relative to obtuse angle point and sharp-pointed angle point, the θ ' of round angle vertex neighborhood inward flange point _lk () value increase ratio is comparatively slow, and the marginal point θ ' in obtuse angle point and sharp comer vertex neighborhood _lk () value increase ratio is comparatively rapid, thus comparatively speaking, the θ ' of circular angle point _lθ ' in (k) value and its neighborhood _lk () mean value relatively.According to this characteristic, carry algorithm according to the marginal point θ ' in angle point neighborhood _lk () value constructs adaptive threshold T _k, shown in (4), if angle point θ ' _lk () value is greater than its local auto-adaptive threshold value, then this angle point is circular angle point, rejects from candidate angular set, if angle point θ ' _lk () value is less than its local auto-adaptive threshold value, be then sharp-pointed angle point or obtuse angle point.

$T_k=x\×\frac{1}{L_1+L_2+1}\underset{i=k-L_1}{\overset{k+L_2}{\mathrm{\Σ}}}{{\mathrm{\θ}}^{,}}_L\left(i\right)---\left(4\right)$

Wherein, x is scale-up factor, L ₁for the length of left neighborhood, L ₂for the length of right neighborhood.Following two key issues existed in once said method are discussed: the selection of Size of Neighborhood and scale-up factor.

3.1 Size of Neighborhood

No matter about selected Size of Neighborhood, angle point left and right sides maximum point can be selected to be boundary, but such contiguous range is too large, be the marginal point θ ' of circular angle point or non-circular angle point, neighborhood both-side ends _l(k) value all close to 1, if these marginal points are taken into account words can make adaptive threshold entirety increase, be unfavorable for whether distinguish candidate angular is circular angle point.Consider that candidate angular is to calculate θ ' by chord length L _lk () value obtains, left and right neighborhood length can be decided to be chord length L, shown in (5), wherein, and Ω (P _k) be angle point P _kcontiguous range, P _k-iwith P _k+ibe respectively angle point P _ki-th pixel in the neighborhood of left and right.

Ω(P _k)＝{P _k-L,...,P _k-1,P _k,P _k+1,...P _k+L}(5)

Experiment shows, above-mentioned contiguous range has good effect, but but fails for contiguous angle point, and the angle point that will be close to that algorithm can be wrong is rejected.In adjacent corners neighborhood of a point, owing to there is other angle points, so its neighborhood Ω (P _k) θ ' of inward flange point _lk () value does not increase progressively always, cause the average theta in this neighborhood ' _lk () value reduces, make angle point θ ' _lk () value is greater than its local auto-adaptive threshold value, thus contiguous angle point is disallowable.So carry algorithm and finally to combine in chord length and neighborhood maximum point to define the contiguous range of angle point, shown in (6):

Ω(P _k)＝{P _{k-min(L,left)},...,P _k-1,P _k,P _k+1,...P _{k+min(L,right)}}(6)

Wherein min (a, b) is for getting the smaller value in a, b, and left is the length to left side maximum point, and right is the length to the right higher value point, i.e. L ₁=min (L, left), L ₂=min (L, right).

3.2 scale-up factor

Definition due to circular angle point is more difficult, is difficult to determine so scale-up factor compares.Be be mathematical without any angle point as a circular object, but for an oval object, if the ratio b/a of its major axis and minor axis is comparatively large, then the angle point at its major axis place is just very flat, can think that this angle point is true angle point.

Experimental design two groups of images test angle point θ ' _laverage theta in (k) value and selected neighborhood ' _lthe ratio relation of (k) value.First group is fillet image, and the ratio (i.e. b/a) of minor axis and major axis is in interval [Isosorbide-5-Nitrae .2], and be spaced apart 0.2, totally 17 images, experimental result is as shown in Fig. 7 (a), and the span of their ratio is [0.96,1].Second group of image is obtuse angle, wedge angle image, and in interval [10 °, 170 °], be spaced apart 10 °, totally 17 images, experimental result is as shown in Fig. 7 (b), and the span of their ratio is [0.58,0.96].

Due in the process that oval angle point increases in b/a value, when ratio is more than 4.2, its characteristic under discretize has been partial to obtuse angle point, so can be considered true angle point, and angle point more than 170 ° is also not too obvious, therefore scale-up factor x value 0.96 herein.

4, false angle point is removed

Because curve is discretize, and there is noise effect, so the θ ' of curved edge point _lk () minimum point can be a lot, wherein major part is false angle point.When constructing angle point local auto-adaptive threshold value and removing circular angle point, although the θ ' of some false angle point own _lk () value is just very little, and the marginal point θ ' in it and neighborhood _lk () value is very nearly the same, thus make its θ ' _lk () value is less than local auto-adaptive threshold value and disallowable, but also have most false angle point still to keep down.

The θ ' of false angle point _lalthough k () value is all less, their codomain is still relatively large, and the threshold value being difficult to selection one fixing distinguishes true angle point and false angle point.Algorithm of carrying removes false angle point accurately in order to set a global threshold, have employed the method for ' many strings ', and true angle point and false angle point can better be distinguished, and method is as follows:

Adopt length to be respectively the string of 10,20,30, obtain respective θ ' respectively such as formula (7) _l(k) value:

${\mathrm{\θ}}_{L_j}^{\′}\left(k\right)=\frac{{\mathrm{\θ}}_{L_j}\left(k\right)}{(L_j-1)\mathrm{\π}},j=1,2,3---\left(7\right)$

Then such as formula (8), the eigenwert Φ (k) that product obtains candidate angular is carried out to three:

$\mathrm{\Φ}\left(k\right)=\underset{j=1}{\overset{3}{\Π}}(1-{\mathrm{\θ}}_{L_j}^{\′}(k\left)\right)---\left(8\right)$

This eigenwert can allow true angle point more distinguish over false angle point, as the θ ' of true angle point _lk () value is respectively 0.2,0.25,0.3, the θ ' of false angle point _lk () value is respectively 0.8,0.85,0.9.Get 1 and deduct θ ' _lk, after (), the value of true angle point is respectively 0.8,0.75,0.7, the value of pseudo-angle point is respectively 0.2,0.15,0.1, and their ratio is respectively 4 times, 5 times, 7 times.After carrying out three's product, eigenwert Φ (k)=0.42 of true angle point, the eigenwert of false angle point is Φ (k)=0.003, and after product, the eigenwert of true angle point is 140 times of false angle point.

As shown in Figure 8, this four width figure is respectively and adopts chord length to be 10,20,30 [the 1-θ ' obtained to Fig. 4 (a) _l(k)] value, and the product of their threes.As can be seen from Fig. 8 (a), 8 (b), 8 (c), true angle point and false angle point are not well distinguished, so a global threshold effectively can not be set remove false angle point, and as can be seen from 8 (d), true angle point and false angle point obtain good differentiation, and making the unified global threshold of setting one remove false angle point becomes possibility.Owing to being when using shorter string to obtain candidate angular before, so contiguous angle point is well distinguished and is remained, the string that now employing three is longer calculates Corner Feature value mistake can't reject contiguous angle point, because now contiguous angle point need not reach maximum value, as long as its eigenwert is greater than global threshold.

Lower surface construction global threshold T rejects false angle point, because obtuse angle point is θ ' in true angle point _lk () value is maximum, so adopt obtuse angle to construct global threshold such as formula shown in (9) herein:

$T={(1-\frac{angle}{180})}^3---\left(9\right)$

Wherein, angle is the obtuse angle of artificial setting, can set different angle values according to different demands, selects 170 ° herein and calculates global threshold T.The angle point being greater than threshold value T is true angle point, and the angle point being less than threshold value T is false angle point, rejects from candidate angular.

Can detect the angle point in image through above step, as shown in Figure 9, Fig. 9 is the net result that original image carries out Corner Detection, and what indicate with ' ' is final angle point.

Claims (1)

1., based on the self-adaptive angular-point detection method that angle is cumulative, specifically comprise the following steps:

(1) adopt canny operator to extract edge, carry out smooth curve by a wicket gaussian kernel;

(2) adopt L value be 6 string carry out the angle accumulated value of edge calculation point, minimalization point is candidate angular;

In step (2), first define the concept that angle is cumulative: the pixel on edge is denoted as P ₁, P ₂..., P _k, P _k+1..., P _n, n is curve up contour point number, for marginal point P each on curve _k, string C _lfrom P _k-L+1p _k+1position starts, and connects string P _k-L+1p _kwith string P _kp _k+1, calculate the angle theta of these two strings _{k, k-L+1}, then string C _lmove to right a pixel to P _k-L+2p _k+2, calculate string P _k-l+2p _kwith string P _kp _k+2angle theta _{k, k-L+2}, by parity of reasoning, until string C _lmove to P _k-1p _k+L-1till position, calculate all angle sums, be marginal point P _kthe angle accumulated value θ at place _l(k);

${\mathrm{\θ}}_L\left(k\right)=\underset{i=k-L+1}{\overset{k-1}{\Σ}}{\mathrm{\θ}}_{k,i}---\left(1\right)$

Then a shorter string C is got _l, L=6, the θ of each pixel on calculated curve _lk () value, due to θ _{k, k-L+j}scope be [0, π], so θ _lk the scope of () is [0, (L-1) π], in order to make experimental data more directly perceived, angle accumulated value being carried out formula (2) conversion and obtaining θ ' _l(k), its scope is [0,1], and minimalization point is incorporated to candidate angular set;

${\mathrm{\θ}}_L^{\′}\left(k\right)=\frac{{\mathrm{\θ}}_L\left(k\right)}{(L-1)\mathrm{\π}}---\left(2\right)$

(3) the local auto-adaptive angle threshold constructing candidate angular removes circular angle point;

In step (3), by the marginal point θ ' in angle point neighborhood _lk () value constructs adaptive threshold T _k, shown in (3), if the θ ' of candidate angular _lwhat k () was greater than adaptive threshold is circular angle point, rejects from candidate angular set;

$T_k=x\×\frac{1}{L_1+L_2+1}\underset{i=k-L_1}{\overset{k+L_2}{\mathrm{\Σ}}}{{\mathrm{\θ}}^{,}}_L\left(i\right)---\left(3\right)$

Wherein, x is scale-up factor, and value is 0.96; L ₁and L ₂be respectively left and right Size of Neighborhood, L ₁=min (L, left), L ₂=min (L, right), L=6, left, right are respectively a P _kcount in edge to left and right limit maximum point;

(4) string adopting L value to be respectively 10,20,30 to candidate angular carrys out the eigenwert of calculated candidate angle point, and constructs global threshold and remove pseudo-angle point, obtains final angle point set;

In step (4), length is adopted to be respectively the string of 10,20,30, such as formula (4), the eigenwert Φ (k) that product obtains candidate angular is carried out to three, then global threshold T is constructed by formula (5), angle is the obtuse angle value of artificial setting, if the eigenwert Φ (k) of candidate angular is less than global threshold, be then false angle point, reject from candidate angular set;

$\mathrm{\Φ}\left(k\right)=\underset{j=1}{\overset{3}{\Π}}(1-{\mathrm{\θ}}_{L_j}^{\′}(k\left)\right)---\left(4\right)$

$T={(1-\frac{angle}{180})}^3---\left(5\right).$