CN103679738A - Image edge detection method based on color radius adjacent domain pixel classification - Google Patents

Abstract

The invention relates to an image edge detection method, in particular to an image edge detection method based on color radius adjacent domain pixel classification and belongs to the technical field of image processing. In the technical scheme, the color image edge detection method based on the color radius adjacent domain pixel classification comprises the following steps of (a) using a Gaussian filter to perform smoothing on color images; (b) adopting an adjacent domain pixel classification method to perform edge treatment on the smoothed color images so as to achieve edge pixel classification of the color images; (c) refining the edges of the color images so as to obtain stable color image edges. The image edge detection method is convenient to operate, the speed of color image edge detection is improved, the edge detection accuracy is high, the adaptability is high, and the stability is reliable.

Description

Method for detecting image edge based on the classification of color radius adjacent domains pixel

Technical field

The present invention relates to a kind of method for detecting image edge, especially a kind of method for detecting image edge based on the classification of color radius adjacent domains pixel, belongs to the technical field that image is processed.

Background technology

Rim detection is an important subject during image is processed, and in computer vision and area of pattern recognition, is widely used.In general, traditional edge detection method comprises three basic steps, is first image pre-service or image filtering, is then image difference and gradient calculation, finally carries out edge extracting.The method of using gradient operator to measure as marginality has had the development of comparative maturity.These edge detection methods mainly can be divided three classes:

1), use difference approximation image function derivative operator, as Roberts operator, Laplace operator, Prewitt operator, Sobel operator, Kirsch operator etc.

2) operator, based on image function second derivative zero crossing, as Marr-Hildreth operator, Canny operator, LoG (Laplacian of Gaussian) operator etc.

3), attempt operator that the parameter model at image function and edge is matched, Haralick and Shapiro did relevant research.

Nearest correlative study has uses Hopfield neural network to carry out the method for rim detection, and Wang Gang etc. have proposed a kind of edge detection method based on sub-pix multifractal.But above-mentioned these methods are all to carry out rim detection for gray level image.

At present, for coloured image, carry out rim detection and mainly contain two kinds of thinkings.Be that coloured image is converted to a gray level image, then use gray-scale Image Edge Detection device to process this width image.This is a kind of many-to-one mapping by color space to the conversion of gray space, the namely conversion from higher dimensional space to lower dimensional space, detected edge accuracy rate can reduce, and meanwhile, in coloured image, significantly border may be lost in gray space.Another kind of thinking is that each color component of coloured image is carried out to rim detection, finally each testing result is carried out to Fusion Edges.The resulting edge of this processing mode is still not accurate enough, and easily ignores the processing to information between color component.Therefore, find a kind of color image edge detection method that can make full use of color image information very meaningful.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, a kind of color image edge detection method based on the classification of color radius adjacent domains pixel is provided, it is easy to operate, the speed of raising to color images edge detection, rim detection precision is high, and strong adaptability is reliable and stable.

According to technical scheme provided by the invention, a kind of color image edge detection method based on the classification of color radius adjacent domains pixel, described color image edge detection method comprises the steps:

A, use Gaussian filter carry out smoothing processing to coloured image;

B, to the coloured image after above-mentioned smoothing processing, utilize adjacent domains pixel sorting technique to carry out edge treated to coloured image, to obtain the edge pixel classification of coloured image;

C, the edge of above-mentioned coloured image is carried out to refinement, to obtain the edge of stable coloured image.

In described step a, use Gaussian filter to carry out when level and smooth coloured image, choosing Size of Neighborhood is 3 * 3, and standard deviation is 0.45.

In described step b, utilize adjacent domains pixel sorting technique to carry out edge treated to coloured image and comprise the steps:

B1, on coloured image, choose a neighborhood, the size of described neighborhood is nb_height*nb_width, and color radius is color_radius, and the threshold value of distinguishing noise class pixel and edge class pixel is noise_edge_t; In described neighborhood, neighborhood territory pixel set is

Set＝{L _i},i＝1,2,…,n

Wherein, Set is neighborhood territory pixel set, L _ifor the pixel in neighborhood, n=nb_height*nb-width; The centre of neighbourhood is:

$\stackrel{\‾}{L}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{L}_i;$

The radius of neighbourhood is:

$R=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left|\right|L_i-\stackrel{\‾}{L}\left|\right|;$

The key words sorting of each pixel in b2, initialization neighborhood, is the key words sorting state of each pixel in neighborhood unfiled, and the initial value of order traversal parameter l oc is 1, and traversal mark mark initial value is 0;

Each pixel in b3, traversal neighborhood element set Set; If loc < is n+1, meanwhile, pixel L _locwhile not being classified, with pixel L _locfor reference point is classified, and forward b4 to; If pixel L _locbe classified, traversal mark mark added up to 1, until mark=n stops the traversal of choosing neighborhood territory pixel S set et to described, and jump to step b7; If L ₁～L _nall be traversed, during loc > n, jump to step b5;

B4, with pixel L _locfor reference point is classified, each pixel in traversal Set, if there is pixel L _knot yet be classified, k gets 1～n, calculating pixel L _locwith pixel L _keuclidean distance, if | L _locl _k|≤color_radius, so pixel L _kbelong to pixel L _locclass, pixel L _locclass in number of pixels add 1, finally obtain with pixel L _locclassification S for reference point _locthe number of middle pixel, is designated as | S _loc|, to traversal parameter l, oc adds up 1, forwards step b3 to;

B5, complete a subseries traversal after, obtain one group of data | S _i|, i=1,2 ..., m, m is once the number of reference point in traversal, gets max{|S _i| corresponding class S _maxfor this time travels through the class marking off, if | S _max|≤noise_edge_t, so mark in such pixel be noise, otherwise the pixel in such is labeled as and is classified, forward afterwards b6 to;

B6, initialization traversal parameter l oc, order traversal parameter l oc=1, forwards step b3 to and continues traversal, until all pixels are all marked as and classify;

B7, to after neighborhood element set Set classification, note classification results is:

Clf_Result＝{S ₁,S ₂,...,S _q}

If | Clf_Result|=1, i.e. q=1, such is 1 region so, does not comprise edge pixel, jumps to step b9; Otherwise, if q >=2 are handled as follows:

Calculate respectively each classification S _v, v=1,2 ..., the class center C of q _vif, note S _v={ L ₁, L ₂..., L _p, so:

$C_v=\frac{1}{p}\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{L}_j$

Calculate each class center C _vto the centre of neighbourhood

distance D _v:

$D_v=\left|\right|C_v-\stackrel{\&OverBar;}{L}\left|\right|$

Calculate distance D _jabsolute value with the difference of radius of neighbourhood R:

d _v＝|D _v-R|

Therefore, obtain one group of range data:

d＝{d ₁,d ₂,...,d _q}

If d _j=max{d}, so j classification S _jin element comprised edge pixel;

B8, to having comprised the class S of edge pixel _j={ L _j1, L _j2..., L _jhin pixel screen; Get classification S _w={ L _w1, L _w2..., L _wk, meet d _w=min{d}, edge pixel L _jx, x=1,2 ..., h need to meet two conditions below:

I, L _jx∈ S _j, and L _jx8 neighborhoods in have pixel L _wy∈ S _w, y=1,2 ... k;

Ii, edge pixel L _jxthere is in one direction edge pixel, be designated as pixel L1, pixel L2, in the direction vertical with this direction, have saltus step pixel, be designated as pixel N1, pixel N2, have

||L-L1||≤color_radius；

||L-L2||≤color_radius；

||L-N1||＞color_radius；

||L-N2||＞color_radius；

To having comprised the class S of edge pixel _jafter screening, obtain edge pixel class S _edge={ L _e1, L _e2..., L _et, described edge pixel class S _edgefor class S _jsubset;

B9, on coloured image, choose next neighborhood, and repeat above-mentioned steps, until whole coloured image is all selected to detection, according to edge pixel class S _edgecomplete the rim detection to coloured image.

Advantage of the present invention: set about from the residing color space of coloured image, directly use chromatic information, carry out image pixel classification in small neighbourhood, according to the classification characteristics of edge pixel, directly mark off edge pixel class.Finally use edge pixel constraint condition edge pixel to screen and obtain desirable image border.This kind of color image edge detection method computation complexity is low, can reach requirement accurately and fast, and has stronger robustness.

Embodiment

Below in conjunction with specific embodiment, the invention will be further described.

For the edge of sense colors image effectively, color image edge detection method of the present invention comprises the steps:

A, use Gaussian filter carry out smoothing processing to coloured image;

In the embodiment of the present invention, by Gaussian filter, coloured image is carried out to smoothing processing, tentatively to remove the noise of coloured image, use Gaussian filter to carry out when level and smooth coloured image, choosing Size of Neighborhood is 3 * 3, and standard deviation is 0.45.

B, to the coloured image after above-mentioned smoothing processing, utilize adjacent domains pixel sorting technique to carry out edge treated to coloured image, to obtain the edge pixel classification of coloured image;

For dim dimension color space (as RGB coloured image dim=3), define color radius C olor_Radius, if in this color space with a C (C ₁, C ₂..., C _dim) be reference point define color, so for the some P (P that belongs to arbitrarily " color C " ₁, P ₂..., P _dim), meet || C-P||≤color_radius (|| the Euclidean distance that C-P|| is point-to-point transmission), the set of the some P of character has formed " color C " like this.Wherein:

$\left|\right|C-P\left|\right|=\sqrt{\underset{i=1}{\overset{\dim }{\mathrm{\&Sigma;}}}{(C_i-P_i)}^2}$

Neighborhood for a nb_height*nb_width, if certain pixel is edge pixel, its " degree of amassing wealth by heavy taxation contribution " minimum to this neighborhood so, namely use color radius C olor_Radius to carry out color classification to the pixel in this neighborhood, edge pixel belongs to certain classification, and the center of this classification is maximum to the absolute value of the distance of the centre of neighbourhood and the difference of the radius of neighbourhood.

Therefore, utilizing adjacent domains pixel sorting technique to carry out edge treated to coloured image comprises the steps:

B1, on coloured image, choose a neighborhood, the size of described neighborhood is nb_height*nb_width, and color radius is color_radius, and the threshold value of distinguishing noise class pixel and edge class pixel is noise_edge_t; In described neighborhood, neighborhood territory pixel set is

Set＝{L _i},i＝1,2,…,n

Wherein, Set is neighborhood territory pixel set, L _ifor the pixel in neighborhood, n=nb_height*nb_width; The centre of neighbourhood is:

$\stackrel{\&OverBar;}{L}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{L}_i;$

The radius of neighbourhood is:

$R=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\left|\right|L_i-\stackrel{\&OverBar;}{L}\left|\right|;$

In the embodiment of the present invention, color radius color_radius, distinguish the size of the threshold value noise_edge_t of noise class pixel and edge class pixel can be as required and concrete color image size select, usually, color radius color_radius can be chosen to be 20, and the threshold value noise_edge_t that distinguishes noise class pixel and edge class pixel can be chosen to be 1 or 2.

The key words sorting of each pixel in b2, initialization neighborhood, is the key words sorting state of each pixel in neighborhood unfiled, and the initial value of order traversal parameter l oc is 1, and traversal mark mark initial value is 0;

Each pixel in b3, traversal neighborhood element set Set; If loc < is n+1, meanwhile, pixel L _locwhile not being classified, with pixel L _locfor reference point is classified, and forward b4 to; If pixel L _locbe classified, traversal mark mark added up to 1, until mark=n stops the traversal of choosing neighborhood territory pixel S set et to described, and jump to step b7; If L ₁～L _nall be traversed, during loc > n, jump to step b5;

B4, with pixel L _locfor reference point is classified, each pixel in traversal Set, if there is pixel L _knot yet be classified, k gets 1～n, calculating pixel L _locwith pixel L _keuclidean distance, if || L _loc-L _k||≤color_radius, so pixel L _kbelong to pixel L _locclass, pixel L _locclass in number of pixels add 1, finally obtain with pixel L _locclassification S for reference point _locthe number of middle pixel, is designated as | S _loc|, to traversal parameter l, oc adds up 1, forwards step b3 to;

B5, complete a subseries traversal after, obtain one group of data | S _i|, i=1,2 ..., m, m is once the number of reference point in traversal, gets max{|S _i| corresponding class S _maxfor this time travels through the class marking off, if | S _max|≤noise_edge_t, so mark in such pixel be noise, otherwise the pixel in such is labeled as and is classified, forward afterwards b6 to; In this step, when | S _max|≤noise_edge_t, so class S _maxin pixel be labeled as noise, otherwise, by class S _maxin pixel be labeled as and classify.

B6, initialization traversal parameter l oc, order traversal parameter l oc=1, forwards step b3 to and continues traversal, until all pixels are all marked as and classify; By step b6, can determine that pixels all in selected neighborhood all can be labeled classification, so that follow-up classification is processed and the operation of screening.

B7, to after neighborhood element set Set classification, note classification results is:

Clf_Result＝{S ₁,S ₂,...,S _q}

If | Clf_Result|=1, i.e. q=1, such is 1 region so, does not comprise edge pixel, jumps to step b9; Otherwise, if q >=2 are handled as follows:

Calculate respectively each classification S _v, v=1,2 ..., the class center C of q _vif, note S _v={ L ₁, L ₂..., L _p, herein, class S _vin comprise p pixel, the value of p is 2 to n, so:

$C_v=\frac{1}{p}\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{L}_j$

Calculate each class center C _vto the centre of neighbourhood

distance D _v:

$D_v=\left|\right|C_v-\stackrel{\&OverBar;}{L}\left|\right|$

Calculate distance D _jabsolute value with the difference of radius of neighbourhood R:

d _v＝|D _v-R|

Therefore, obtain one group of range data:

d＝{d ₁,d ₂,...,d _q}

If d _j=max{d}, so j classification S _jin element comprised edge pixel;

In the embodiment of the present invention, by the class in classification results Clf_Result is calculated, can access one group of range data.

B8, to having comprised the class S of edge pixel _j={ L _j1, L _j2..., L _jhin pixel screen; Get classification S _w={ L _w1, L _w2..., L _wk, meet d _w=min{d}, edge pixel L _jx, x=1,2 ..., h need to meet two conditions below:

I, L _jx∈ S _j, and L _jx8 neighborhoods in have pixel L _wy∈ S _w, y=1,2 ... k;

Ii, edge pixel L _jxthere is in one direction edge pixel, be designated as pixel L1, pixel L2, in the direction vertical with this direction, have saltus step pixel, be designated as pixel N1, pixel N2, have

||L-L1||≤color_radius；

||L-L2||≤color_radius；

||L-N1||＞color_radius；

||L-N2||＞color_radius；

To having comprised the class S of edge pixel _jafter screening, obtain edge pixel class S _edge={ L _e1, L _e2..., L _et, described edge pixel class S _edgefor class S _jsubset;

B9, on coloured image, choose next neighborhood, and repeat above-mentioned steps, until whole coloured image is all selected to detection, according to edge pixel class S _edgecomplete the rim detection to coloured image.

C, the edge of above-mentioned coloured image is carried out to refinement, to obtain the edge of stable coloured image.

In the embodiment of the present invention, the coloured image after detecting by edge carries out refinement, to obtain stable edge, thinning method wherein can reference literature " Lam; L., Seong-Whan Lee, and Ching Y.Suen; " Thinning Methodologies-A Comprehensive Survey, explanation in " IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol14, No.9; September1992, page879 ".

The color image edge detection method that the present invention is based on the classification of color radius adjacent domains pixel, algorithm is simple, is easy to realize.And relative Canny edge detection operator, carries out rim detection and has superiority having added the coloured image of salt-pepper noise.

The present invention sets about from the residing color space of coloured image, directly uses color image information, carries out image pixel classification in small neighbourhood, according to the classification characteristics of edge pixel, directly marks off edge pixel class.Finally use edge pixel constraint condition edge pixel to screen and obtain desirable image border.This kind of color image edge detection method computation complexity is low, can reach requirement accurately and fast, and has stronger robustness.

Claims (3)

1. the color image edge detection method based on the classification of color radius adjacent domains pixel, is characterized in that, described color image edge detection method comprises the steps:

(a), use Gaussian filter to carry out smoothing processing to coloured image;

(b), to the coloured image after above-mentioned smoothing processing, utilize adjacent domains pixel sorting technique to carry out edge treated to coloured image, to obtain the edge pixel classification of coloured image;

(c), the edge of above-mentioned coloured image is carried out to refinement, to obtain the edge of stable coloured image.

2. the color image edge detection method based on color radius adjacent domains pixel classification according to claim 1, it is characterized in that, in described step (a), use Gaussian filter to carry out when level and smooth coloured image, choosing Size of Neighborhood is 3 * 3, and standard deviation is 0.45.

3. the color image edge detection method based on the classification of color radius adjacent domains pixel according to claim 1, is characterized in that, in described step (b), utilizes adjacent domains pixel sorting technique to carry out edge treated to coloured image and comprises the steps:

(b1), on coloured image, choose a neighborhood, the size of described neighborhood is nb_height*nb_width, color radius is color_radius, the threshold value of distinguishing noise class pixel and edge class pixel is noise_edge_t; In described neighborhood, neighborhood territory pixel set is

Set＝{L _i},i＝1,2,…,n

Wherein, Set is neighborhood territory pixel set, L _ifor the pixel in neighborhood, n=nb_height*nb_width; The centre of neighbourhood is:

$\stackrel{\&OverBar;}{L}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{L}_i;$

The radius of neighbourhood is:

$R=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\left|\right|L_i-\stackrel{\&OverBar;}{L}\left|\right|;$

(b2), the key words sorting of each pixel in initialization neighborhood, the key words sorting state of each pixel in neighborhood is unfiled, the initial value of order traversal parameter l oc is 1, traversal mark mark initial value is 0;

(b3), each pixel in traversal neighborhood element set Set; If loc < is n+1, meanwhile, pixel L _locwhile not being classified, with pixel L _locfor reference point is classified, and forward (b4) to; If pixel L _locbe classified, traversal mark mark added up to 1, until mark=n stops the traversal of choosing neighborhood territory pixel S set et to described, and jump to step (b7); If L ₁～L _nall be traversed, during loc > n, jump to step (b5);

(b4), with pixel L _locfor reference point is classified, each pixel in traversal Set, if there is pixel L _knot yet be classified, k gets 1～n, calculating pixel L _locwith pixel L _keuclidean distance, if || L _loc-L _k||≤color_radius, so pixel L _kbelong to pixel L _locclass, pixel L _locclass in number of pixels add 1, finally obtain with pixel L _locclassification S for reference point _locthe number of middle pixel, is designated as | S _loc|, to traversal parameter l, oc adds up 1, forwards step (b3) to;

(b5), complete a subseries traversal after, obtain one group of data | S _i|, i=1,2 ..., m, m is once the number of reference point in traversal, gets max{|S _i| corresponding class S _maxfor this time travels through the class marking off, if | S _max|≤noise_edge_t, so mark in such pixel be noise, otherwise the pixel in such is labeled as and is classified, forward afterwards (b6) to;

(b6), initialization traversal parameter l oc, order traversal parameter l oc=1, forwards step (b3) to and continues traversal, until all pixels are all marked as and classify;

(b7), to after neighborhood element set Set classification, note classification results is:

Clf_Result＝{S ₁,S ₂,...,S _q}

If | Clf_Result|=1, i.e. q=1, such is 1 region so, does not comprise edge pixel, jumps to step (b9); Otherwise, if q >=2 are handled as follows:

Calculate respectively each classification S _v, v=1,2 ..., the class center C of q _vif, note S _v={ L ₁, L ₂..., L _p, so:

$C_v=\frac{1}{p}\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{L}_j$

Calculate each class center C _vto the centre of neighbourhood

distance D _v:

$D_v=\left|\right|C_v-\stackrel{\&OverBar;}{L}\left|\right|$

Calculate distance D _jabsolute value with the difference of radius of neighbourhood R:

d _v＝|D _v-R|

Therefore, obtain one group of range data:

d＝{d ₁,d ₂,...,d _q}

If d _j=max{d}, so j classification S _jin element comprised edge pixel;

(b8), to having comprised the class S of edge pixel _j={ L _j1, L _j2..., L _jhin pixel screen; Get classification S _w={ L _w1, L _w2..., L _wk, meet d _w=min{d}, edge pixel L _jx, x=1,2 ..., h need to meet two conditions below:

(i), L _jx∈ S _j, and L _jx8 neighborhoods in have pixel L _wy∈ S _w, y=1,2 ... k;

(ii), edge pixel L _jxthere is in one direction edge pixel, be designated as pixel L1, pixel L2, in the direction vertical with this direction, have saltus step pixel, be designated as pixel N1, pixel N2, have

||L-L1||≤color_radius；

||L-L2||≤color_radius；

||L-N1||＞color_radius；

||L-N2||＞color_radius；

To having comprised the class S of edge pixel _jafter screening, obtain edge pixel class S _edge={ L _e1, L _e2..., L _et, described edge pixel class S _edgefor class S _jsubset;

(b9), on coloured image, choose next neighborhood, and repeat above-mentioned steps, until whole coloured image is all selected to detection, according to edge pixel class S _edgecomplete the rim detection to coloured image.