CN102331425A - Textile defect detection method based on defect enhancement - Google Patents

Abstract

The invention discloses a textile defect detection method based on defect enhancement. The textile defect detection method is characterized by comprising the following steps of: acquiring various types of defected parts and non-defected parts of a textile by using a charge coupled device (CCD) camera so as to obtain a defected sample and a non-defected sample respectively, and finishing image acquisition; enhancing the acquired image by using the similarity among local modes, wherein as the similarity of the modes is measured in a main component space, partial noise can be eliminated and the similarity of the modes can be measured more precisely; and texture information of the image of the textile can be kept better; and performing three-layer Haar wavelet decomposition on the enhanced image of the textile, cutting a defect on a wavelet sub band with the maximum defect energy distribution by using a threshold value cutting method, outputting a cutting result and finally finishing defect detection. By the method, the texture information of the image of the textile can be kept better and differences between the defected parts and the non-defected parts can be reflected remarkably; and therefore, the correct rate of detection can be increased.

Description

Method for detecting textile defect based on the defective enhancing

Technical field

The present invention relates to a kind of method for detecting textile defect, particularly be applied to the detection of digital image processing techniques to the defective that causes by mechanical disorder and yarn problem.

Background technology

Defects detection is the key link that industrial textile is produced.In textile industry, textile defect has kind more than 50, and most of defectives are caused that by mechanical disorder and yarn problem this type defective can be divided into soiled yarn, spider web, cracked ends and latitude, shyer and six kinds of defectives of loose yarn.The textile defect of manual type detects subjective experience, notice and the judgment that effect depends critically upon the tester.In the Modern Textile article commercial production, textile defect detects automatically and is progressively replacing manual detection.

Existing textile defect Automatic Measurement Technique based on Digital Image Processing and pattern-recognition is divided into two parts usually: study part and test section.At first, zero defect training sample and defective training sample are learnt, obtained priori, obtain reference vector.Then, carry out sample classification, realize that textile defect detects according to the similarity degree between reference vector and the test sample book.Concrete study and testing process generally include: IMAQ, graphical analysis, eigenwert are extracted and defects detection.When the utilization image processing techniques detects textile automatically; Owing to receive the influence of noise and the material of fabric own in the image pattern acquisition process; The texture structure of sample is smudgy, is unfavorable for the extraction of characteristics of image, makes that existing detection method verification and measurement ratio is low.So be necessary the defective of textile images is strengthened, reach and reduce noise effect, improve the purpose of verification and measurement ratio.

In nearly ten years, the enhancing of the texture of image has received extensive concern, can be divided into picture contrast enhancing and picture structure and strengthen.It mainly is the histogram distribution of expanded images that picture contrast strengthens.Method commonly used has histogram equalization, gradient algorithm, based on texture Enhancement Method of Lifting Wavelet or the like.The stretched grey level histogram of image of these methods distributes, and has strengthened the brightness of image, but to a little less than the capability of restraining noise, only is fit to cause because of Luminance Distribution is uneven the enhancing of the image that blurs.The enhancing of picture structure mainly is to highlight the texture of image and edge.Up-to-date method has the texture Enhancement Method based on differential, and this method has been given prominence to the texture information at edge, but to noise-sensitive, is fit to not have the image of noise.Enhancement Method based on multiple dimensioned morphology reconstruct has not only strengthened contrast, has also kept the correlativity of the gray level in continuous zone, and this method is fit to the uneven image of some Luminance Distribution, is not suitable for noise image.Grain details enhancing based on consciousness comprises three parts: Clipped Median Filter, Sober filtering and contrast on border strengthen, and this method has overcome the defective of crossing enhancing and weak enhancing in the gradient enhancing.Above Enhancement Method is the enhancing to general texture image, is not suitable for the enhancing of the textile images with abundant, regular periods texture structure.There is multi-dimension texture to strengthen and based on the textile images Enhancement Method of discrete gaussian filter to the Enhancement Method of this particular image.The former is proposed in nineteen ninety-five by Joachim Weickert, mainly is to improve the method for anisotropy diffusion, applies to then in the textile images, and this method has kept the continuity of image texture.To be Bao Xiaomin proposed in 2005 the latter, and this method is the level and smooth texture of image when removing noise.The texture structure of textile images is complicated, more than two kinds of methods all be difficult to the texture information of maintenance textile images.

Summary of the invention

The present invention is for avoiding above-mentioned existing in prior technology weak point, proposing a kind of method for detecting textile defect that strengthens based on defective.To improve anti-noise ability, can keep the texture information of textile images well, highlight the difference of defect part and non-defect part, be convenient to correctly cutting apart of defects detection part, detect accuracy thereby improve.

Technical solution problem of the present invention adopts following technical scheme:

The characteristics that the present invention is based on the method for detecting textile defect of defective enhancing are to be undertaken by following process:

A, IMAQ: utilize ccd video camera to gather the various types of defect parts and the non-defect part of textile, obtain defect sample and zero defect sample respectively;

B, defective strengthen: utilize the weighted value of neighborhood territory pixel point to represent central pixel point, eliminate the noise that produces in the imaging process; Weight w (i, j) similarity by central pixel point i and its neighborhood territory pixel point j determines; The similarity of pixel i and pixel j depends on neighborhood characteristics vector v (N _i) and v (N _j) similarity degree; Said neighborhood territory pixel point is meant with pixel i to be the search window S at center _iInterior pixel j, N _iBe meant with pixel i be the center be the neighborhood territory pixel piece of size with similar window; Said defective strengthens and is undertaken by following mode:

At first with neighborhood characteristics vector v (N _i) project to the principal component space and obtain new proper vector v [d] (N _i); The utilization Euclidean distance is by the similarity degree of formula (1) tolerance texture in said principal component space; Similarity degree according to texture calculates weights by formula (2) again; At last the neighborhood territory pixel v (j) of pixel v (i) is carried out the weighting back pixel value μ (i) that is enhanced by formula (3), the textile images μ after being enhanced after all pixels of image v are strengthened respectively as stated above;

||v[d](N _i)-v[d](N _j)|| (1)

$w(i,j)=\frac{1}{M\left(i\right)}{e}^{-\frac{\left|\right|v\left[d\right]\left(N_i\right)-v\left[d\right]\left(N_j\right)\left|\right|}{h^2}}---\left(2\right)$

In the formula (2) $M\left(i\right)=\underset{j\∈S_i}{\mathrm{\Σ}}{e}^{-\frac{\left|\right|v\left[d\right]\left(N_i\right)-v\left[d\right]\left(N_j\right)\left|\right|}{h^2}}$ I=1,2......k

$\mathrm{\μ}\left(i\right)=\underset{j\∈S_i}{\mathrm{\Σ}}w(i,j)v\left(j\right)---\left(3\right)$

H wherein is a damped expoential, and k is the pixel number of image;

C, defects detection: the textile images to after strengthening is carried out wavelet transformation, on the wavelet sub-band of defective energy distribution maximum, uses threshold segmentation method to carry out defective and cuts apart, and the output segmentation result is finally accomplished defects detection.

The characteristics that the present invention is based on the method for detecting textile defect of defective enhancing also are:

Defective among the said step b strengthen be with the neighborhood characteristics vector projection to the principal component space, d component calculating similarity before selecting, d value are taken as 6,7......49; Damped expoential h is taken as 0.5 δ ², δ ²Be noise variance; According to pixels the meter search window is taken as 21 * 21, and similar window is taken as 7 * 7.

Weighted model among the said step b is to adopt the arest neighbors weighted model, and said weighted model is that a kind of maximum weights all pixels in the search window are composed the weighted type to the weights of central pixel point.

The wavelet function that among the said step c textile images after strengthening is carried out wavelet transformation adopts the Haar small echo, and the textile images after the said enhancing is carried out three layers of wavelet decomposition.

Compared with present technology, beneficial effect of the present invention is embodied in:

1, before detecting, at first textile images has been carried out the defective enhancing among the present invention; Eliminated the noise effect of imaging system and fabric itself; Highlight the difference of defect part and non-defect part, be convenient to correctly cutting apart of defect part, detect accuracy thereby improve.

2, the present invention has adopted a kind of new defective enhancement algorithms, utilizes the similarity between local mode that the textile images sample is strengthened.Principal component spatial measure pattern similarity property can not only eliminate a part of noise and also can be more the similarity of accurate measurement pattern, thereby better keep the texture information of textile images.

3, textile images has periodic texture structure, and generation of defects makes local texture that distortion take place.Wavelet transformation can effectively be distinguished the texture of this distortion among the present invention on different yardsticks and direction, better is partitioned into defect part.

Description of drawings

Fig. 1 is the inventive method synoptic diagram;

Fig. 2 is defective enhancement principle figure in the inventive method;

Fig. 3 is a defects detection schematic diagram in the inventive method.

Embodiment

Present embodiment Applied Digital image processing techniques detects the defective by mechanical disorder or the caused textile of yarn problem, and process is following:

Referring to Fig. 1, IMAQ: utilize ccd video camera to gather the various types of defect parts and the non-defect part of textile, obtain defect sample and zero defect sample respectively;

Referring to Fig. 1, Fig. 2, it is to utilize the weighted value of neighborhood territory pixel point to represent central pixel point that defective strengthens, and eliminates additive noise; Weight w (i, j) similarity by central pixel point i and its neighborhood territory pixel point j determines; The similarity of pixel i and pixel j depends on neighborhood characteristics vector v (N _i) and v (N _j) similarity degree; The neighborhood territory pixel point is meant with pixel i to be the search window S at center _iInterior pixel j, N _iBe meant with pixel i be the center be the neighborhood territory pixel piece of size with similar window; Defective strengthens and is undertaken by following mode:

At first with neighborhood characteristics vector v (N _i) project to the principal component space and obtain new proper vector v [d] (N _i); The utilization Euclidean distance is by the similarity degree of formula (1) tolerance texture in the principal component space; Similarity degree according to texture calculates weights by formula (2) again; At last the neighborhood territory pixel v (j) of pixel v (i) is carried out the weighting back pixel value μ (i) that is enhanced by formula (3), the textile images μ after being enhanced after all pixels of image v are strengthened respectively as stated above;

||v[d](N _i)-v[d](N _j)|| (1)

$w(i,j)=\frac{1}{M\left(i\right)}{e}^{-\frac{\left|\right|v\left[d\right]\left(N_i\right)-v\left[d\right]\left(N_j\right)\left|\right|}{h^2}}---\left(2\right)$

In the formula (2) $M\left(i\right)=\underset{j\∈S_i}{\mathrm{\Σ}}{e}^{-\frac{\left|\right|v\left[d\right]\left(N_i\right)-v\left[d\right]\left(N_j\right)\left|\right|}{h^2}}$ I=1,2......k

$\mathrm{\μ}\left(i\right)=\underset{j\∈S_i}{\mathrm{\Σ}}w(i,j)v\left(j\right)---\left(3\right)$

H wherein is a damped expoential, and k is the pixel number of image;

In the present embodiment with the neighborhood characteristics vector projection to the principal component space, d component calculating similarity before selecting, d value are taken as 6,7......49; According to pixels the meter search window is taken as 21 * 21, and similar window is taken as 7 * 7, and damped expoential h is taken as 0.5 δ ², δ ²Be noise variance; Weighted model is to adopt the arest neighbors weighted model, and the arest neighbors weighted model is that a kind of maximum weights all pixels in the search window are composed the weighted type to the weights of central pixel point.

Because the textile images of obtaining in the automatic detection imaging system of textile receives the interference of noise, fabric material and tiny knitting wool etc., has had a strong impact on testing result.Can remove the interference that causes by above factor in the image through the defective enhancing in the present embodiment, recover the intrinsic tissue texture information of textile, highlight the difference of defect part and non-defect part simultaneously.Textile images has the texture information of regular periods, has a large amount of block of pixels with parallel pattern, and present embodiment utilizes the similarity of these local modes to recover texture information, has made full use of the architectural feature of textile images.

Because textile images receives the influence of noise and structural complexity thereof; Be not independently between each neighborhood characteristics vector; And the assumed condition of utilization Euclidean distance calculating similarity degree is that each neighborhood characteristics vector is separate; Therefore, strengthen two point defects below the textile images existence in pixel space: 1, can not accurately measure pattern similarity property.2, a little less than the noise resisting ability of similarity.Principal component analysis is a kind of typical statistical analysis technique, is decomposed into image information component and noise component to noise image, and has removed the correlativity between each component.In the present embodiment, can be more in the principal component space accurate measurement pattern similarity improving anti-noise ability, to keep the texture information of textile images well, thereby reaches the purpose that highlights the textile images defect area.The dimension d in principal component space can select adjustment through the analysis of separability between defect sample and non-defect sample, thereby makes the separability between defective enhanced results maximization defective and non-defective, further improves the accuracy rate of defects detection.Concrete embodiment is that the d value is taken as 6,7......49, type of obtaining separability result respectively, and the pairing d value of best class separability is exactly final value.

Referring to Fig. 1, Fig. 3, defects detection is that the textile images after strengthening is carried out wavelet transformation, on the wavelet sub-band of defective energy distribution maximum, uses threshold segmentation method to carry out defective and cuts apart, and the output segmentation result is finally accomplished defects detection.

The wavelet function that in the present embodiment textile images after strengthening is carried out wavelet transformation adopts the Haar small echo, and the textile images after strengthening is carried out three layers of wavelet decomposition.

For the maximum wavelet sub-band 1～N of the energy distribution that obtains all kinds of defectives and the segmentation threshold of all kinds of defectives, N class defective training sample is learnt, obtain priori.Because each type defect image all can be maximum in the energy distribution on some wavelet sub-bands, the energy distribution on other subbands can be ignored.So test sample image is carried out wavelet transformation, and on 1～N small echo wavelet sub-band, cuts apart, always have a kind of wavelet sub-band and be fit to sample image.Synthetic to 1～N segmentation result, finally export testing result.

Claims (4)

1. the method for detecting textile defect that strengthens based on defective is characterized in that being undertaken by following process:

A, IMAQ: utilize ccd video camera to gather the various types of defect parts and the non-defect part of textile, obtain defect sample and zero defect sample respectively;

B, defective strengthen: utilize the weighted value of neighborhood territory pixel point to represent central pixel point, eliminate the noise that produces in the imaging process; Weight w (i, j) similarity by central pixel point i and its neighborhood territory pixel point j determines; The similarity of pixel i and pixel j depends on neighborhood characteristics vector v (N _i) and v (N _j) similarity degree; Said neighborhood territory pixel point is meant with pixel i to be the search window S at center _iInterior pixel j, N _iBe meant with pixel i be the center be the neighborhood territory pixel piece of size with similar window; Said defective strengthens and is undertaken by following mode:

At first with neighborhood characteristics vector v (N _i) project to the principal component space and obtain new proper vector v [d] (N _i); The utilization Euclidean distance is by the similarity degree of formula (1) tolerance texture in said principal component space; Similarity degree according to texture calculates weights by formula (2) again; At last the neighborhood territory pixel v (j) of pixel v (i) is carried out the weighting back pixel value μ (i) that is enhanced by formula (3), the textile images μ after being enhanced after all pixels of image v are strengthened respectively as stated above;

||v[d](N _i)-v[d](N _j)|| (1)

$w(i,j)=\frac{1}{M\left(i\right)}{e}^{-\frac{\left|\right|v\left[d\right]\left(N_i\right)-v\left[d\right]\left(N_j\right)\left|\right|}{h^2}}---\left(2\right)$

In the formula (2) $M\left(i\right)=\underset{j\∈S_i}{\mathrm{\Σ}}{e}^{-\frac{\left|\right|v\left[d\right]\left(N_i\right)-v\left[d\right]\left(N_j\right)\left|\right|}{h^2}}$ I=1,2......k

$\mathrm{\μ}\left(i\right)=\underset{j\∈S_i}{\mathrm{\Σ}}w(i,j)v\left(j\right)---\left(3\right)$

H wherein is a damped expoential, and k is the pixel number of image;

C, defects detection: the textile images to after strengthening is carried out wavelet transformation, on the wavelet sub-band of defective energy distribution maximum, uses threshold segmentation method to carry out defective and cuts apart, and the output segmentation result is finally accomplished defects detection.

2. according to the described method for detecting textile defect that strengthens based on defective of right 1, it is characterized in that defective among the said step b strengthen be with the neighborhood characteristics vector projection to the principal component space, d component calculating similarity before selecting, d value are taken as 6,7......49; Damped expoential h is taken as 0.5 δ ², δ ²Be noise variance; According to pixels count, search window is taken as 21 * 21, and similar window is taken as 7 * 7.

3. according to the right 1 described method for detecting textile defect that strengthens based on defective; It is characterized in that the weighted model among the said step b is to adopt the arest neighbors weighted model, said weighted model is that a kind of maximum weights all pixels in the search window are composed the weighted type to the weights of central pixel point.

4. the method for detecting textile defect that based on defective strengthen described according to right 1; It is characterized in that among the said step c that the wavelet function that the textile images after strengthening is carried out wavelet transformation adopts the Haar small echo, and the textile images after the said enhancing is carried out three layers of wavelet decomposition.

Images