CN105261003A - Defect point detection method on basis of self structure of fabric - Google Patents

Abstract

The invention discloses a defect point detection method on the basis of a self structure of a fabric. The method comprises the following concrete detection steps of: preprocessing a collected grey cloth fabric image; then, enhancing a feature value of a defect point image; next, performing denoising processing on the enhanced defect point image; segmenting the defect point image by a maximum between-class variance threshold segmentation method; and finally, performing fabric defect point detection by an area filtering method. The method has the advantages that the operation starts from the self structure of the fabric; a vein element template is used for enhancing a defect point region; a zero-mean image is used for completing the enhancement on the defect point image; regions of a gray scale mutation defect point, a structure mutation defect point and a mixed defect point can be effectively enhanced; the limitation that a standard defect-point-free image needs to be trained by a conventional method can be avoided; and the accuracy, the reliability and the applicability of the defect point detection method are improved.

Description

A kind of method of carrying out defect detection based on fabric self structure

Technical field

The invention belongs to textile appearance defect detection method and technology field, be specifically related to a kind of method of carrying out defect detection based on fabric self structure.

Background technology

It is the important content that modern enterprise is produced that textile quality detects, and the defect detection of current fabric mainly contains manual detection and automatically detects two kinds.Artificial vision check false drop rate and loss higher, can only find the fabric defects of 70 ~ 80%, testing result by larger, the detection speed of subjective impact slowly, high in cost of production defect.Existing automatic testing method sums up the detection method, the detection method based on frequency-domain analysis and the detection method three major types based on model that are mainly divided into Corpus--based Method.But because fabric defects is of a great variety, come in every shape, add the complicacy of cloth textured image itself, existing detection method is often only adapted to the fault detecting some particular type, cannot detect all faults, bad adaptability;

Fabric is interwoven according to certain organization rule by longitude and latitude two groups of yarns, and warp, broadwise all have certain periodicity.Periodic texture can be thought to be made up of many texture primitives (elementary cell of texture) close to each other, braiding mutually.Normally all follow certain specific rule without the pixel grey scale Distribution value in texture primitive each in the textile image of fault, similarity between different texture primitives is strong, and the appearance of fault, no matter be grey degree type fault or structural type fault, grey value profile rule in texture primitive must be destroyed.Therefore fabric defects detection can be started with from fabric self structure study completely.

Summary of the invention

The object of this invention is to provide a kind of method of carrying out defect detection based on fabric self structure, effectively can detect gray scale saltant type fault, structural mutation type fault and Combination fault, improve the accuracy of defect detection method, reliability and applicability.

The technical solution adopted in the present invention is, a kind of method of carrying out defect detection based on fabric self structure, and concrete detecting step is:

Step 1: pre-service is carried out to the greige goods fabric image gathered;

Step 2: defect image eigenwert strengthens;

Step 3: structure average image, carries out denoising to the defect image after strengthening;

Step 4: adopt maximum between-cluster variance thresholding method to split defect image;

Step 5: the detection being carried out fabric defects by the method for area filtering.

Feature of the present invention is also,

In step 1, Image semantic classification comprises image enhaucament and image denoising.

The process of image enhaucament is: subwindow image being divided into p × p size of non-overlapping copies, and p is the multiple of 8; Calculate the average MV of each subwindow gray scale _i,j, that is:

${\mathrm{MV}}_{i,j}=\underset{m=0}{\overset{p}{\Σ}}\underset{n=0}{\overset{p}{\Σ}}{G}_{i+m,j+n}$

In formula, MV _i,jfor the gray average of image i capable j row pixel, G _i,jfor the gray-scale value of image i capable j row pixel;

Utilize interpolation formula (1) to carry out bilinear interpolation, extend to original size, obtain average image:

$\begin{array}{c}{G}_{p*i+m,p*j+n}={\mathrm{MV}}_{i,j}+\frac{m}{p}*\left({\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j}\right)+\frac{n}{p}\left({\mathrm{MV}}_{i,j+1}-{\mathrm{MV}}_{i,j}\right)+\\ \frac{m*n}{p\×p}\left({\mathrm{MV}}_{i+1,j+1}+{\mathrm{MV}}_{i,j}-{\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j+1}\right)\end{array}---\left(1\right)$

In formula, m, n ∈ [0, p-1];

Finally, utilize respective pixel gray scale difference between original image and average image to construct required zero-mean image, namely complete image enhaucament.

The process of image denoising is: suppose I=f (x, y) for there being the image of N × N number of pixel, its smoothing operator r=h (i, j), size is K × K, the image I ' that the width obtained after utilizing formula (2) smoothing process is new=g (x, y), realize image denoising, that is:

$g(x,y)=\frac{1}{M}\underset{m=-K/2}{\overset{K/2}{\Σ}}\underset{n=-K/2}{\overset{K/2}{\Σ}}f(x+m,y+n)*h(m,n)---\left(2\right)$

In formula: x, y ∈ [0, N-1]; M is weights sum in operator.

In step 2, the process that defect image eigenwert strengthens is:

Step 2.1: adopt autocorrelation function computed image in the horizontal direction with the auto-correlation function value of vertical direction, that is:

$C_{x,0}=\frac{\frac{1}{M*(N-x)}\underset{i=1}{\overset{N-x}{\Σ}}\underset{j=1}{\overset{M}{\Σ}}{G}_{i,j}*G_{i+x,j}}{\frac{1}{M*N}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M}{\Σ}}{G_{i,j}}^2}---\left(3\right)$

$C_{0,y}=\frac{\frac{1}{N*(M-y)}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M-y}{\Σ}}{G}_{i,j}*G_{i,j+y}}{\frac{1}{M*N}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M}{\Σ}}{G_{i,j}}^2}---\left(4\right)$

In formula, C _{x, 0}for the auto-correlation function value of horizontal direction; C _{0, y}for the auto-correlation function value of vertical direction; M*N represents the size of image, G _i,jrepresent the gray-scale value of pixel, cycle T _x, T _yfor the size of texture primitive;

Step 2.2: textile image to be detected is divided into T _x× T _ysubwindow, utilizes following formula to calculate the gray-scale value of pixel in son calculating primitive template:

$M_{i,j}=\frac{1}{n}\underset{k=1}{\overset{n}{\Σ}}{W}_{i,j}^k$

In formula, M _i,jrepresent the gray-scale value of pixel in primitive template, represent the gray-scale value of pixel in a kth subwindow, n represents the number of subwindow;

Step 2.3: textile image to be detected is divided into T again _x× T _ysubwindow, by formula S ^k _i,j=| W ^k _i,j-M _i,j| calculate the difference S between subwindow and primitive template ^k _i,j, utilize formula (5) to all S ^k _i,jafter carrying out the process of linear gauge generalized, obtain fault and strengthen image, that is:

$S_{i,j}^{\′}=\frac{\left({S^k}_{i,j}-Min\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)\right)*255}{Max\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)-Min\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)}---\left(5\right)$

In formula, S ' _i,jfor fault strengthens the gray-scale value of image pixel; Max (), Min () are respectively maximal value and minimum value function.

In step 3, the process of image denoising process is:

Step 2 gained image is divided into the subwindow of p × p, calculates the average MV of each window gray scale _i,j: ${\mathrm{MV}}_{i,j}=\underset{m=0}{\overset{p}{\Σ}}\underset{n=0}{\overset{p}{\Σ}}{G}_{i+m,j+m},$

In formula, G _i,jfor the gray-scale value of image pixel, utilize interpolation formula (6) to carry out bilinear interpolation, extend to original size, obtain average image:

$\begin{array}{c}{G}_{p*i+m,p*j+n}={\mathrm{MV}}_{i,j}+\frac{m}{p}*\left({\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j}\right)+\frac{n}{p}\left({\mathrm{MV}}_{i,j+1}-{\mathrm{MV}}_{i,j}\right)+\\ \frac{m*n}{p\×p}\left({\mathrm{MV}}_{i+1,j+1}+{\mathrm{MV}}_{i,j}-{\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j+1}\right)\end{array}---\left(6\right)$

In formula, m, n ∈ [0, p-1].

In step 4, the concrete steps that defect image carries out splitting are:

Image level gray scale is made to be L, to each gray-scale value x, p _xrepresent the frequency that x occurs, now segmentation threshold is t, then gray scale is divided into two classes: C1=(0,1 ..., t); C2=(t+1, t+2 ..., L); The probability that then each class occurs is: $w_1\left(t\right)=\underset{x=1}{\overset{t}{\Σ}}{p}_x$ With $w_2\left(t\right)=\underset{x=t+1}{\overset{L-1}{\Σ}}{p}_x=1-w_1\left(t\right);$ The average gray of each class is: $u_1=\frac{u\left(t\right)}{w_1\left(t\right)}$ With $u_2=\frac{u-u\left(t\right)}{1-w_1\left(t\right)},$ Wherein $u\left(t\right)=\underset{x=1}{\overset{t}{\Σ}}{\mathrm{xp}}_x,u=\underset{x=1}{\overset{L-1}{\Σ}}{\mathrm{xp}}_x;$ Then inter-class variance is:

σ ²(t)＝w ₁(u-u ₁) ²+w ₂(u ₂-u) ²＝w ₁w ₂(u ₂-u ₁) ²(7)

Make σ ²t () gets the t of maximal value, be exactly the optimal threshold of segmentation object and background, and what gray-scale value was greater than t is target area to be split.

In step 5, the detailed process of fabric defects detection is:

Adopt the connected region of the bianry image of eight connectivity method after step 4 is split being searched for non-zero pixels value, the connected region label running into first non-zero pixels value is 1, and the number recording this area pixel point is M1; By that analogy, to the last a non-zero pixels region labeling is N, and the number of area pixel is MN; Here Mi is exactly the area of i-th connected region; If the textile image that to be detected, the area of connected regions all after Iamge Segmentation is all less than some threshold values, then this width image does not just have fault; Otherwise if there is the connected region that area is greater than threshold value after Iamge Segmentation, then this width image comprises fault, the position of fault, size just can be greater than the connected region position of threshold value by these and area is determined.

The invention has the beneficial effects as follows, the present invention starts with from fabric self structure, utilize texture primitive template to strengthen the method for defect regions, and adopt zero-mean image to complete to strengthen defect image, effectively can strengthen gray scale mutability fault, structural mutation fault and Combination defect regions, and the restriction that classic method can be avoided need to train without defect image standard, improve the accuracy of defect detection method, reliability and applicability.

Accompanying drawing explanation

Fig. 1 is a kind of process flow diagram carrying out defect detection method based on fabric self structure of the present invention;

Fig. 2 is the periodicity figure of textile image texture;

Fig. 3 is the design sketch utilizing the inventive method to carry out crapand defect detection;

Fig. 4 is the design sketch utilizing the inventive method to carry out flotation line defect detection;

Fig. 5 is the design sketch utilizing the inventive method to carry out broken yarn defect detection;

Fig. 6 is the design sketch utilizing the inventive method to carry out watermark stain defect detection.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

The present invention a kind ofly carries out the flow process of defect detection method as shown in Figure 1 based on fabric self structure, and concrete steps are as follows:

Step 1: Image semantic classification:

Image semantic classification mainly comprises enhancing and the image denoising of image;

The present invention utilizes zero-mean image enchancing method.First image is divided subwindow one by one, in order to piece image being divided into an integer subwindow, the size of subwindow should be the approximate number of image size; Subwindow can not be too small in addition, should be larger than the size of fault, thus reach the object of fault enhancing.Comprehensive above 2 points, the size of the subwindow that the present embodiment adopts is 16 × 16.The size of subwindow calculates the average MV of each window gray scale after determining _i,j, g in formula _i,jrepresent the gray-scale value of pixel, and carry out down-sampling; Utilize interpolation formula (1) to carry out bilinear interpolation, extend to original size, obtain average image:

$\begin{array}{c}{G}_{16*i+m,16*j+n}={\mathrm{MV}}_{i,j}+\frac{m}{16}*\left({\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j}\right)+\frac{n}{16}\left({\mathrm{MV}}_{i,j+1}-{\mathrm{MV}}_{i,j}\right)+\\ \frac{m*n}{256}\left({\mathrm{MV}}_{i+1,j+1}+{\mathrm{MV}}_{i,j}-{\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j+1}\right)\end{array}---\left(1\right)$

Finally, original image is utilized just to construct required zero-mean image with the difference of the average image obtained by interpolation.Zero-mean image can not only increase the intensity contrast of image, and due to the average of each subwindow be zero, therefore, it is possible to effectively solve the light and shade difference problem caused because of illumination.

Image denoising adopts the mean filter of 3 × 3.Mean filter adopts neighborhood averaging, assuming that a width has the image I=f (x of N × N number of pixel, y), smoothing operator r=h (i, j), its size is K × K, the image I ' that the width obtained after smoothing processing is new=g (x, y), so there is formula (2) to set up:

$g(x,y)=\frac{1}{M}\underset{m=-K/2}{\overset{K/2}{\Σ}}\underset{n=-K/2}{\overset{K/2}{\Σ}}f(x+m,y+n)*h(m,n)---\left(2\right)$

In formula: x, y ∈ [0, N-1]; M is weights sum in operator.

Step 2: the enhancing of defect image eigenwert, specifically implements according to the following steps

Step 2.1: determine texture primitive size.Determine texture primitive size.Fabric is interwoven according to certain organization rule by through weft yarn, and warp, broadwise all have certain periodicity, and textile image belongs to the orderly texture image of rule.Autocorrelator trace presents good periodicity, and vertical and horizontal one-period forms a texture primitive, as shown in Figure 2.The size of texture primitive adopts auto-correlation (Auto-correlation) function usually, and the autocorrelation function computing method of image level, vertical direction are:

$C_{0,y}=\frac{\frac{1}{N*(M-y)}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M-y}{\Σ}}{G}_{i,j}*G_{i,j+y}}{\frac{1}{M*N}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M}{\Σ}}{G_{i,j}}^2}---\left(4\right)$

In formula, M*N represents the size of image, G _i,jrepresent the gray-scale value of pixel, C _{x, 0}, C _{0, y}the auto-correlation function value of expression level, vertical direction, cycle T _x, T _yfor the size of texture primitive.

Step 2.2: the calculating of texture primitive template, after the size of texture primitive is determined, then calculates texture primitive template.Method is: original image is divided into T _x× T _ysubwindow, grey scale pixel value corresponding in all subwindows is averaged, and computing formula is: in formula, M _i,jrepresent the gray-scale value of pixel in primitive template, represent the gray-scale value of pixel in a kth subwindow, n represents the number of subwindow.

Step 2.3: the enhancing of defect regions, the automatic detection of fabric defects needs constantly to slacken background, i.e. normal texture area information, simultaneously outstanding defect regions information, so that finally defect detection out.Fabric defect image to be detected is divided into T again _x× T _ysubwindow, by formula S ^k _i,j=| W ^k _i,j-M _i,j| calculate the difference between subwindow and primitive template, then to all S ^k _i,jcarry out the process of linear gauge generalized, disposal route is:

$S_{i,j}^{\′}=\frac{\left({S^k}_{i,j}-Min\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)\right)*255}{Max\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)-Min\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)}---\left(5\right)$

In formula, Max (), Min () are respectively maximal value and minimum value function, and we obtain fault enhancing image by this method.

Step 3: structure average image, detailed process is:

Defect image after enhancing, owing to having a large amount of high frequency noises, directly can not carry out binary segmentation, is carried out the method for stress release treatment by structure average image.The building method of average image is: image is divided into the subwindow of 8 × 8, calculates the average MV of each window gray scale _i,j, and carry out down-sampling, then carry out bilinear interpolation, extend to original size, interpolation formula is such as formula shown in (6):

$\begin{array}{c}{G}_{8*i+m,8*j+n}={\mathrm{MV}}_{i,j}+\frac{m}{8}*\left({\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j}\right)+\frac{n}{8}\left({\mathrm{MV}}_{i,j+1}-{\mathrm{MV}}_{i,j}\right)+\\ \frac{m*n}{64}\left({\mathrm{MV}}_{i+1,j+1}+{\mathrm{MV}}_{i,j}-{\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j+1}\right)\end{array}---\left(6\right)$

Average image can eliminate high frequency noise information effectively, is convenient to next step defect segmentation, the average image of structure.

Step 4: the segmentation of defect image: adopt maximum between-cluster variance thresholding method method, specifically implement according to the following steps:

Maximum between-cluster variance thresholding method is also referred to as Otsu method, first based on the probability of happening of each gray level of histogram calculation, and with thresholding variables t, gray level is divided into two classes, then variance within clusters and the inter-class variance of each class is obtained, choose and make inter-class variance maximum, the minimum t of variance within clusters is as optimal threshold.

If image has L level gray scale, to each gray-scale value x, p _xrepresent the frequency that x occurs, now segmentation threshold is t, then gray scale is divided into two classes: C1=(0,1 ..., t); C2=(t+1, t+2 ..., L).The probability that then each class occurs is: with $w_2\left(t\right)=\underset{x=t+1}{\overset{L-1}{\Σ}}{p}_x=1-w_1\left(t\right);$ The average gray of each class is: $u_1=\frac{u\left(t\right)}{w_1\left(t\right)}$ With $u_2=\frac{u-u\left(t\right)}{1-w_1\left(t\right)},$ Wherein then inter-class variance can be defined by formula:

σ ²(t)＝w ₁(u-u ₁) ²+w ₂(u ₂-u) ²＝w ₁w ₂(u ₂-u ₁) ²(7)

Make σ ²t () gets the t of maximal value, be exactly the optimal threshold of segmentation object and background, and what gray-scale value was greater than t is target area to be split.

Step 5: the detection of fault: some very little white portion of the image after segmentation is not fault, needs to adopt area filtering that these very little white portions are removed.The computing method of connected region area bianry image after singulation adopt eight connectivity method search for the connected region of non-zero pixels value, and the connected region label running into first non-zero pixels value is 1, and the number recording this area pixel point is M1.By that analogy, to the last a non-zero pixels region labeling is N, and the number of area pixel is MN.Here Mi is exactly the connected region area of requirement.Area be less than certain threshold value we just do not think that it is fault.

Principle of the present invention is: defect detection is actually a Texture Segmentation and identifying, because the texture structure at fault place is different from normal fabric in fabric, therefore, it is possible to they are detected.Fabric is interwoven according to certain organization rule by through weft yarn, and warp, broadwise all have certain periodicity, and therefore textile image belongs to the orderly texture image of rule.As shown in Figure 2, autocorrelator trace presents good periodicity, and vertical and horizontal one-period forms a texture primitive.In normal fabric image, in each texture primitive, the gray-scale value of respective pixel relatively, and defect regions, no matter be that gray scale saltant type fault is as greasy dirt, broken hole etc., or structural mutation type fault as and warp, loose warp etc., the gray-scale value of respective pixel and the larger change of the existence of normal fabric in texture primitive.Defect detection method based on the analysis of fabric self structure puts forward based on this principle just.First it comprise zero-mean image procossing and mean filter to the pre-service of textile image to be detected, then autocorrelation function determination texture primitive size is calculated, calculate texture primitive template, texture primitive template is utilized to realize the enhancing of defect regions information, then average image is constructed, utilize Otsu method to realize defect image segmentation, finally utilize area filtering to determine whether textile image to be detected contains the detection of fault and fault.Based on fabric self structure analyze defect detection method to structural mutation type fault (as crapand), as shown in Figure 3; Mixed type fault (as flotation line and broken yarn), is shown in Figure 4 and 5; Gray scale mutability fault (as watermark stain), is shown in Fig. 6; This method all has good segmentation effect to above-mentioned fault as can be seen from Figure, and in Fig. 3, crapand fault is split well and detects; Flotation line segmentation in Fig. 4 is also relatively good with Detection results; The broken yarn fault being exposed at fabric face in Fig. 5 is better detected; Watermark stain in same Fig. 6 is all effectively split and is detected.

Zero-mean image method of the present invention not only complete well to defect image strengthen requirement, and due to zero-mean image acquiring method simpler than histogram equalization method many, time efficiency is high; Start with from fabric self structure, utilize texture primitive template to strengthen the method for defect regions, effectively can strengthen gray scale mutability fault, structural mutation fault and Combination defect regions, and the restriction that classic method need be trained without defect image standard can be avoided; According to textile image texture own characteristic, start with from this very important visual feature of periodicity of image texture, propose the fabric defect detection method analyzed based on fabric self structure, to all kinds of fault, all there is good Detection results, improve the accuracy of defect detection method, reliability and applicability; Whole algorithm of the present invention does not have complex calculations, and time efficiency is high, can meet the needs that fabric defects real-time online detects.

Claims (8)

1. carry out a method for defect detection based on fabric self structure, it is characterized in that, concrete detecting step is:

Step 1: pre-service is carried out to the greige goods fabric image gathered;

Step 2: defect image eigenwert strengthens;

Step 3: structure average image, carries out denoising to the defect image after strengthening;

Step 4: adopt maximum between-cluster variance thresholding method to split defect image;

Step 5: the detection being carried out fabric defects by the method for area filtering.

2. a kind of method of carrying out defect detection based on fabric self structure according to claim 1, is characterized in that, in step 1, Image semantic classification comprises image enhaucament and image denoising.

3. a kind of method of carrying out defect detection based on fabric self structure according to claim 2, is characterized in that, the process of described image enhaucament is: subwindow image being divided into p × p size of non-overlapping copies, and p is the multiple of 8; Calculate the average MV of each subwindow gray scale _i,j, that is:

${\mathrm{MV}}_{i,j}=\underset{m=0}{\overset{p}{\Σ}}\underset{n=0}{\overset{p}{\Σ}}{G}_{i+m,j+n},$

In formula, MV _i,jfor the gray average of image i capable j row pixel, G _i,jfor the gray-scale value of image i capable j row pixel;

Utilize interpolation formula (1) to carry out bilinear interpolation, extend to original size, obtain average image:

$\begin{array}{c}{G}_{p*i+m,p*j+n}={\mathrm{MV}}_{i,j}+\frac{m}{p}*\left({\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j}\right)+\frac{n}{p}\left({\mathrm{MV}}_{i,j+1}-{\mathrm{MV}}_{i,j}\right)+\\ \frac{m*n}{p\×p}\left({\mathrm{MV}}_{i+1,j+1}+{\mathrm{MV}}_{i,j}-{\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j+1}\right)\end{array},---\left(1\right)$

In formula, m, n ∈ [0, p-1];

Finally, utilize respective pixel gray scale difference between original image and average image to construct required zero-mean image, namely complete image enhaucament.

4. a kind of method of carrying out defect detection based on fabric self structure according to claim 2, it is characterized in that, the process of described image denoising is: suppose that I=f (x, y) is for there being the image of N × N number of pixel, its smoothing operator r=h (i, j), size is K × K, the image I ' that the width obtained after utilizing formula (2) smoothing process is new=g (x, y), realize image denoising, that is:

$g(x,y)=\frac{1}{M}\underset{m=-K/2}{\overset{K/2}{\Σ}}\underset{n=-K/2}{\overset{K/2}{\Σ}}f(x+m,y+n)*h(m,n),---\left(2\right)$

In formula: x, y ∈ [0, N-1]; M is weights sum in operator.

5. a kind of method of carrying out defect detection based on fabric self structure according to claim 1, is characterized in that, in step 2, the process that defect image eigenwert strengthens is:

Step 2.1: adopt autocorrelation function computed image in the horizontal direction with the auto-correlation function value of vertical direction, that is:

$C_{x,0}=\frac{\frac{1}{M*\left(N-x\right)}\underset{i=1}{\overset{N-x}{\Σ}}\underset{j=1}{\overset{M}{\Σ}}{G}_{i,j}*G_{i+x,j}}{\frac{1}{M*N}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M}{\Σ}}{G_{i,j}}^2},---\left(3\right)$

$C_{0,y}=\frac{\frac{1}{N*(M-y)}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M-y}{\Σ}}{G}_{i,j}*G_{i,j+y}}{\frac{1}{M*N}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M}{\Σ}}{G_{i,j}}^2},---\left(4\right)$

In formula, C _{x, 0}for the auto-correlation function value of horizontal direction; C _{0, y}for the auto-correlation function value of vertical direction; M*N represents the size of image, G _i,jrepresent the gray-scale value of pixel, cycle T _x, T _yfor the size of texture primitive;

Step 2.2: textile image to be detected is divided into T _x× T _ysubwindow, utilizes following formula to calculate the gray-scale value of pixel in son calculating primitive template:

$M_{i,j}=\frac{1}{n}\underset{k=1}{\overset{n}{\Σ}}{W}_{i,j}^k,$

In formula, M _i,jrepresent the gray-scale value of pixel in primitive template, represent the gray-scale value of pixel in a kth subwindow, n represents the number of subwindow;

Step 2.3: textile image to be detected is divided into T again _x× T _ysubwindow, by formula S ^k _i,j=| W ^k _i,j-M _i,j| calculate the difference S between subwindow and primitive template ^k _i,j, utilize formula (5) to all S ^k _i,jafter carrying out the process of linear gauge generalized, obtain fault and strengthen image, that is:

$S_{i,j}^{\′}=\frac{\left({S^k}_{i,j}-Min\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)\right)*255}{Max\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)-Min\left({S^1}_{i,j},{S^2}_{i,j},...,{S^n}_{i,j}\right)},---\left(5\right)$

In formula, S ' _i,jfor fault strengthens the gray-scale value of image pixel; Max (), Min () are respectively maximal value and minimum value function.

6. a kind of method of carrying out defect detection based on fabric self structure according to claim 1, is characterized in that, in step 3, the process of image denoising process is:

Step 2 gained image is divided into the subwindow of p × p, calculates the average MV of each window gray scale _i,j:

${\mathrm{MV}}_{i,j}=\underset{m=0}{\overset{p}{\Σ}}\underset{n=0}{\overset{p}{\Σ}}{G}_{i+m,j+n},$

In formula, G _i,jfor the gray-scale value of image pixel;

Utilize interpolation formula (6) to carry out bilinear interpolation, extend to original size, obtain average image:

$\begin{array}{c}{G}_{p*i+m,p*j+n}={\mathrm{MV}}_{i,j}+\frac{m}{p}*\left({\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j}\right)+\frac{n}{p}\left({\mathrm{MV}}_{i,j+1}-{\mathrm{MV}}_{i,j}\right)+\\ \frac{m*n}{p\×p}\left({\mathrm{MV}}_{i+1,j+1}+{\mathrm{MV}}_{i,j}-{\mathrm{MV}}_{i+1,j}-{\mathrm{MV}}_{i,j+1}\right)\end{array},---\left(6\right)$

In formula, m, n ∈ [0, p-1].

7. a kind of method of carrying out defect detection based on fabric self structure according to claim 1, is characterized in that, in step 4, the concrete steps that defect image carries out splitting are:

Image level gray scale is made to be L, to each gray-scale value x, p _xrepresent the frequency that x occurs, now segmentation threshold is t, then gray scale is divided into two classes: C1=(0,1 ..., t); C2=(t+1, t+2 ..., L); The probability that then each class occurs is: $w_1\left(t\right)=\underset{x=1}{\overset{t}{\Σ}}{p}_x$ With $w_2\left(t\right)=\underset{x=t+1}{\overset{L-1}{\Σ}}{p}_x=1-w_1\left(t\right);$ The average gray of each class is: $u_1=\frac{u\left(t\right)}{w_1\left(t\right)}$ With $u_2=\frac{u-u\left(t\right)}{1-w_1\left(t\right)},$ Wherein $u\left(t\right)=\underset{x=1}{\overset{t}{\Σ}}{\mathrm{xp}}_x,u=\underset{x=1}{\overset{L-1}{\Σ}}{\mathrm{xp}}_x;$ Then inter-class variance is:

σ ²(t)＝w ₁(u-u ₁) ²+w ₂(u ₂-u) ²＝w ₁w ₂(u ₂-u ₁) ²，(7)

Make σ ²t () gets the t of maximal value, be exactly the optimal threshold of segmentation object and background, and what gray-scale value was greater than t is target area to be split.

8. a kind of method of carrying out defect detection based on fabric self structure according to claim 1, is characterized in that, in step 5, the detailed process of fabric defects detection is:

Adopt the connected region of the bianry image of eight connectivity method after step 4 is split being searched for non-zero pixels value, the connected region label running into first non-zero pixels value is 1, and the number recording this area pixel point is M1; By that analogy, to the last a non-zero pixels region labeling is N, and the number of area pixel is MN; Here Mi is exactly the area of i-th connected region; If the textile image that to be detected, the area of connected regions all after Iamge Segmentation is all less than some threshold values, then this width image does not just have fault; Otherwise if there is the connected region that area is greater than threshold value after Iamge Segmentation, then this width image comprises fault, the position of fault, size just can be greater than the connected region position of threshold value by these and area is determined.