CN104751474A - Cascade quick image defect segmentation method - Google Patents

Abstract

The invention relates to a cascade quick image defect segmentation method. The method includes: preprocessing an image to acquire global threshold information, and performing rough image segmentation based on a global adaptive threshold; acquiring fine image segmentation through priori knowledge and a level set method of local image information; eliminating evident error candidate defects and error alarms by means of size and shape analysis, and performing target separation on the basis of a parameterized mathematical morphology method. the problems such that non-uniformity of local and global thresholds leads to inaccurate image segmentation are solved by means of a stratified strategy; features of defects, such as shape and size, are introduced to a level set energy function; the priori knowledge of the defect shape is fully utilized, the problem of initial dependence is solved, and the method is well applicable to image defect segmentation having shape features; by introducing the cascade image processing architecture and discarding most non-defective images from the most front, unbalance of data distribution is greatly relieved, and defect segmentation speed and precision is greatly improved.

Description

A kind of tandem type rapid image Method of Defect Segmentation

Technical field

The present invention relates to a kind of image processing techniques, particularly a kind of tandem type rapid image Method of Defect Segmentation.

Background technology

From the uneven image of gray scale, be partitioned into defect target is a difficulty and important technology, also be the committed step of subsequent detection and classification, introduce the mankind and can have very large advantage to the segmentation of defect image to the priori of various defect, particularly industrial products surface abnormalities detects and defect inspection.Common way is the image partition method from adopting overall adaptive threshold, such as Otsu method determines optimum intensity slicing threshold value by maximizing inter-class variance and minimizing variance within clusters, but gray scale is uneven, and to make local threshold choose incorrect, is difficult to obtain satisfied segmentation result; On the other hand, the method involved by great mass of data is to the process poor real of defect, and accuracy is not high yet, and its reason is the calculating that costs a lot of money need not on zero defect image to be processed, and therefore these detection methods cannot be applied to actual real-time system.

Through finding existing literature search, the thought of series connection or integrated (boosted) has obtained and has used widely in classification, the people such as Viola propose the boosting method be made up of Multilayer Classifier tandem first, be applied to fast face detection, eye detection and Gender Classification etc., obtain good effect, their thought is that the multiple similar Weak Classifier of general employing is combined into strong classifier, therefore has certain limitation.Do not have documents and materials to be incorporated in Iamge Segmentation by tandem type thinking at present, the information yet not utilizing image different in different partitioning algorithms, such as the information of local and the overall situation, in the segmentation of defect, do not introduce the priori of the mankind.

Summary of the invention

The present invention be directed to and be partitioned into the not high problem of the time-consuming accuracy of defect from the uneven image of gray scale, propose a kind of tandem type rapid image Method of Defect Segmentation, based on Iamge Segmentation, image is divided into specifically, peculiar property region, and obtain some significant target or structures in image.This invention is for the deficiency of conventional images cutting techniques, not only make use of the global information of image, and make use of local message and the priori of image, compensate for now methodical deficiency, good application has been carried out to the image deflects segmentation with shape facility.And in the first step of Iamge Segmentation, eliminate most of zero defect image, substantially increase speed and the accuracy of segmentation.

Technical scheme of the present invention is: a kind of tandem type rapid image Method of Defect Segmentation, the defective image of band is carried out coarse to meticulous image procossing, eliminate the false alarm that strong noise produces step by step, first after the first step carries out pre-service to image, obtain global threshold information, adopt the coarse image based on overall adaptive threshold to split; Then second step utilizes the Level Set Method of priori and local image information to obtain meticulous Iamge Segmentation; Last 3rd step uses size and dimension analysis to eliminate manifest error candidate defect and false alarm, carries out target separation based on parametrization Mathematical Morphology Method.

The described first step adjustment image intensity, contrast, then after adopting 4*4 medium filtering to carry out denoising Processing, use the segmentation of Otsu method coarse image, determine optimum intensity slicing threshold value by maximizing inter-class variance and minimizing variance within clusters, complete the coarse image segmentation of overall adaptive threshold.

Described second step first uses Mathematical Morphology Method to abate the noise, and obtains the curve closed, and the initial value split using this closed curve as precise image, then utilize the Level Set Method comprising priori to obtain meticulous Iamge Segmentation.

The curve that described acquisition is closed adopts area filling method to eliminate cavity and the noncoherent boundary of area-of-interest, then the circular configuration Mathematical Morphology Method of 5 pixel diameters is used to carry out opening and closing operation, filter out very thin border and random noise region, when gray scale is uneven, in order to eliminating error alarm with extract shape facility accurately, have employed Level Set Method refining based on Chan – Vese model from the coarse corresponding subgraph curve split, obtain sectionally smooth object edge.

The shape item with priori is introduced in described meticulous Iamge Segmentation, refining is with the curve of the overall situation with local geometry information, specifically: on the basis of Chan – Vese model, proposing following energy functional, obtaining optimum segmentation by solving this energy functional:

$F(C,c_I,c_o)={\mathrm{\λ}}_1\underset{\mathrm{\Ω}}{\∫}{\left|\right|\widehat{C-C}\left|\right|}^2\mathrm{ds}+{\mathrm{\λ}}_2(\underset{\mathrm{inside}\left(C\right)}{\∫}{(u(x,y)-c_I)}^2\mathrm{dxdy}+\underset{\mathrm{outside}\left(C\right)}{\∫}{(u(x,y)-c_o)}^2\mathrm{dxdy})$

Wherein u (x, y) is the gray-scale value of image pixel (x, y), c _ithe average gray of outline line C inside, c _obe the average gray of outline line C outside, Section 1 is shape item, reflects outline line C and shape function zero level collection between difference, be shape function, latter two is area item, λ ₁and λ ₂be the normal number of equilibrium configuration and area item weight, ds is the arc infinitesimal in the curvilinear integral on the Ω of region;

Described shape function by carrying out shape modeling to same defects, collecting a large amount of same defects and forming a redundant data training set, generating one based on training set by principal component analysis (PCA) PCA new data set:

the average of training concentration curve, X _pcaeigencoefficient vector, shape vector or variation eigenmode, in matrix W _pin p major component is sorted, in practice, only from training data, ask for first principal component, modeling carried out to the maximum variation of shape facility, i.e. p=1.

Described 3rd step employs two parameterized structural elements, a circular configuration element being radius r and changing, another is the line structure element of length l and angle θ change, its width w=1, the crack of arbitrarily angled and size, joint and hole can be separated, adopt structural element B (x) opening operation to set A to be expressed as A ο B (x), be defined as follows:

(1) for the circular configuration element of radius r change,

(2) for the line structure element of length l and angle θ change,

${\mathrm{\φ}}_{\mathrm{open}}\left(l\right)=\underset{i=1,...,18}{\mathrm{Max}}\left\{\mathrm{\φ}(l,{\mathrm{\θ}}_i)\right\}$

Here angle θ _ibetween 0 ° to 180 ° at interval of 10 ° of values once, suitable parameter l and r is chosen according to the length range in crack and the radius in hole.

Beneficial effect of the present invention is: tandem type rapid image Method of Defect Segmentation of the present invention, prior art is compared, select to have most the feature of resolution characteristic by maximizing Pasteur's distance between class, error in classification can be estimated approx, flawless image can be got rid of fast, cost more calculates in minority candidate target image or region, therefore the efficiency of whole system is greatly improved, under pipeline processing mode, the requirement of application in real time can be met, and there is very high verification and measurement ratio, less false positive, noise robust, the features such as efficiency is high, whole cutting procedure is made to have very high precision.

Accompanying drawing explanation

Fig. 1 is total Organization Chart of the layered image segmentation of tandem type Fast image segmentation method of the present invention;

Fig. 2 is that the present invention is applied to the cutting procedure image with two quasi-representative defects and finally splits image;

Fig. 3 is the comparison diagram of image partition method of the present invention and different images dividing method.

Embodiment

The technological means realized to make the present invention, creation characteristic, reaching object and effect is easy to understand, below in conjunction with concrete diagram, setting forth the present invention further.

Total Organization Chart of the layered image segmentation of tandem type Fast image segmentation method as shown in Figure 1.The present embodiment can think the image processing process from coarse to meticulous, eliminates the false alarm that strong noise produces step by step.Whole technical process is divided into three steps: the first step is split based on the coarse image of overall adaptive threshold; Second step is split based on the level set precise image of priori; 3rd step is separated based on the detection target of parametrization Mathematical Morphology Method.Specific as follows:

The first step: to the defective gray level image of band, in order to improve image definition, first Image semantic classification is carried out, adjustment image intensity, contrast etc., then 4*4 medium filtering is adopted to carry out denoising Processing, then the segmentation of Otsu method coarse image is used herein, require under the defect situation that the least possible omission is possible, obtain candidate defect, Otsu is a kind of effectively based on the overall adaptive threshold method of discriminatory analysis, it determines optimum intensity slicing threshold value by maximizing inter-class variance and minimizing variance within clusters, but this method is to noise and target size sensitivity, therefore boundary alignment mistake and alarm by mistake can be produced in coarse image segmentation, as shown in Figure 2, (a) original image, image after (b) Ostu segmentation, the image of (c) mixing Roughen Edges, d () has the fine segmentation image of CV level set red curve, e gray-scale map that () splits by Level Set Method, f () has the binary map of little noise after using circular configuration element (r=1mm) opening operation, noiseless binary map after (g) utilization circular configuration element (r=2mm) opening operation, h () uses the binary map after line structure element (l=15mm) opening operation, i () uses the binary map after line structure element (l=6mm) opening operation, the result figure of (j) (g)-(i).

Second step: first use Mathematical Morphology Method to abate the noise, obtain the curve closed, and the initial value split using this closed curve as precise image, then utilize the Level Set Method comprising priori to obtain meticulous Iamge Segmentation.The present invention adopts area filling method to eliminate cavity and the noncoherent boundary of area-of-interest, for successive image process forms closed curve.Then, use the circular configuration Mathematical Morphology Method of 5 pixel diameters to carry out opening and closing operation, filter out very thin border and random noise region, these borders and noise may be that little fragment or coloured background cause.

When gray scale is uneven, alarm and false edges by mistake can be produced.In order to eliminating error alarm with extract shape facility accurately, adopt the border in each region of image local gamma characteristic refining.There is employed herein Level Set Method refining based on Chan – Vese model from the coarse corresponding subgraph curve split, obtain sectionally smooth object edge.

The described shape item introduced in precise image segmentation with priori, refining is with the curve of the overall situation with local geometry information, specifically: on the basis of CV model, we have proposed following energy functional, obtaining optimum segmentation by solving this energy functional:

$F(C,c_I,c_o)={\mathrm{\λ}}_1\underset{\mathrm{\Ω}}{\∫}{\left|\right|\widehat{C-C}\left|\right|}^2\mathrm{ds}+{\mathrm{\λ}}_2(\underset{\mathrm{inside}\left(C\right)}{\∫}{(u(x,y)-c_I)}^2\mathrm{dxdy}+\underset{\mathrm{outside}\left(C\right)}{\∫}{(u(x,y)-c_o)}^2\mathrm{dxdy})$

Here u (x, y) is the gray-scale value of image pixel (x, y), c _ithe average gray of outline line C inside, c _obe the average gray of outline line C outside, Section 1 is shape item, reflects outline line C and shape function zero level collection between difference, be shape function, latter two is area item, λ ₁and λ ₂be the normal number of equilibrium configuration and area item weight, wherein ds is the arc infinitesimal in curvilinear integral on the Ω of region.

In order to obtain described shape function need to carry out shape modeling to same defects, collect a large amount of same defects and form a redundant data training set, then utilize this training set, adopt principal component analysis (PCA) (PCA) method to remove the redundant information of symbolic measurement, by covariance matrix Σ=MM ^tthe proper vector of/n obtains the major component of symbolic measurement, and M is a matrix here, and its column vector is n training symbol distance function the row of composition.Proper vector { e _j, namely major component, the main transformer of a corresponding n point set divides direction.Therefore one is generated based on training set by PCA new data set:

Here the average of training concentration curve, X _pcaeigencoefficient vector, shape vector or variation eigenmode, in matrix W _pin p major component is sorted.In practice, only ask for first principal component here from training data, carries out modeling, i.e. p=1 to the maximum variation of shape facility.

The comparison diagram of image partition method of the present invention and different images dividing method as shown in Figure 3, (k) Sobel rim detection; (l) Ostu ' s and morphological method; (m) SUSAN method; N () is based on the morphological segment (r=1) of Sobel; O () is based on the morphological method (r=1) of SUSAN; (p) dividing method result of the present invention.

3rd step: dissimilar defect has different geometric configuratioies and size, we use simple size and dimension analysis to eliminate manifest error candidate defect and false alarm, use the Mathematical Morphology Method of different structural elements to be separated different type flaws.The present invention adopts parametrization morphological method to be separated difform defect, the crack of such as wire or cut, the through hole of round shape, the circular hole of ring-type.In order to extract the geometric properties of each defect and remove larger background noise, present invention uses two parameterized structural elements, a circular configuration element being radius r and changing, another is the line structure element of length l and angle θ change, its width w=1, can be separated the crack of arbitrarily angled and size, joint and hole.

Adopt structural element B (x) opening operation to set A to be expressed as A ο B (x), be defined as follows:

(1) for the circular configuration element of radius r change,

(2) for the line structure element of length l and angle θ change,

${\mathrm{\φ}}_{\mathrm{open}}\left(l\right)=\underset{i=1,...,18}{\mathrm{Max}}\left\{\mathrm{\φ}(l,{\mathrm{\θ}}_i)\right\}$

Here angle θ _ibetween 0 ° to 180 ° at interval of 10 ° of values once, suitable parameter l and r is chosen according to the length range in crack and the radius in hole.

Claims (6)

1. a tandem type rapid image Method of Defect Segmentation, it is characterized in that, the defective image of band is carried out coarse to meticulous image procossing, eliminate the false alarm that strong noise produces step by step, first after the first step carries out pre-service to image, obtain global threshold information, adopt the coarse image based on overall adaptive threshold to split; Then second step utilizes the Level Set Method of priori and local image information to obtain meticulous Iamge Segmentation; Last 3rd step uses size and dimension analysis to eliminate manifest error candidate defect and false alarm, carries out target separation based on parametrization Mathematical Morphology Method.

2. tandem type rapid image Method of Defect Segmentation according to claim 1, it is characterized in that, the described first step adjustment image intensity, contrast, then after adopting 4*4 medium filtering to carry out denoising Processing, use the segmentation of Otsu method coarse image, determine optimum intensity slicing threshold value by maximizing inter-class variance and minimizing variance within clusters, complete the coarse image segmentation of overall adaptive threshold.

3. tandem type rapid image Method of Defect Segmentation according to claim 1, it is characterized in that, described second step first uses Mathematical Morphology Method to abate the noise, obtain the curve closed, and the initial value split using this closed curve as precise image, then utilize the Level Set Method comprising priori to obtain meticulous Iamge Segmentation.

4. tandem type rapid image Method of Defect Segmentation according to claim 3, it is characterized in that, the curve that described acquisition is closed adopts area filling method to eliminate cavity and the noncoherent boundary of area-of-interest, then the circular configuration Mathematical Morphology Method of 5 pixel diameters is used to carry out opening and closing operation, filter out very thin border and random noise region, when gray scale is uneven, in order to eliminating error alarm with extract shape facility accurately, have employed Level Set Method refining based on Chan – Vese model from the coarse corresponding subgraph curve split, obtain sectionally smooth object edge.

5. tandem type rapid image Method of Defect Segmentation according to claim 4, it is characterized in that, the shape item with priori is introduced in described meticulous Iamge Segmentation, refining is with the curve of the overall situation with local geometry information, specifically: on the basis of Chan – Vese model, proposing following energy functional, obtaining optimum segmentation by solving this energy functional:

$F(C,c_I,c_o)={\mathrm{\λ}}_1\underset{\mathrm{\Ω}}{\∫}{\left|\right|\widehat{C-C\left|\right|}}^2\mathrm{ds}+{\mathrm{\λ}}_2(\underset{\mathrm{inside}\left(C\right)}{\∫}{(u(x,y)-c_I)}^2\mathrm{dxdy}+\underset{\mathrm{outside}\left(C\right)}{\∫}{(u(x,y)-c_o)}^2\mathrm{dxdy})$

Wherein u (x, y) is the gray-scale value of image pixel (x, y), c _ithe average gray of outline line C inside, c _obe the average gray of outline line C outside, Section 1 is shape item, reflects outline line C and shape function zero level collection between difference, be shape function, latter two is area item, λ ₁and λ ₂be the normal number of equilibrium configuration and area item weight, ds is the arc infinitesimal in the curvilinear integral on the Ω of region;

Described shape function by carrying out shape modeling to same defects, collecting a large amount of same defects and forming a redundant data training set, generating one based on training set by principal component analysis (PCA) PCA new data set:

the average of training concentration curve, X _pcaeigencoefficient vector, shape vector or variation eigenmode, in matrix W _pin p major component is sorted, in practice, only from training data, ask for first principal component, modeling carried out to the maximum variation of shape facility, i.e. p=1.

6. tandem type rapid image Method of Defect Segmentation according to claim 1, it is characterized in that, described 3rd step employs two parameterized structural elements, a circular configuration element being radius r and changing, another is the line structure element of length l and angle θ change, and its width w=1 can be separated the crack of arbitrarily angled and size, joint and hole, adopt structural element B (x) opening operation to set A to be expressed as A o B (x), be defined as follows:

(1) for the circular configuration element of radius r change,

(2) for the line structure element of length l and angle θ change,

${\mathrm{\φ}}_{\mathrm{open}}\left(l\right)=\underset{i=1,...,18}{\mathrm{Max}}\left\{\mathrm{\φ}(l,{\mathrm{\θ}}_i)\right\}$

Here angle θ _ibetween 0 ° to 180 ° at interval of 10 ° of values once, suitable parameter l and r is chosen according to the length range in crack and the radius in hole.