CN103926255A - Method for detecting surface defects of cloth based on wavelet neural network - Google Patents

Abstract

The invention provides an online visual method for detecting surface defects of cloth. The detecting method is characterized in that the information such as width and direction of the surface textures of the cloth can be effectively extracted by combing Gabor filter and the wavelet neural network, the optimal solution of the same kind of cloth is obtained, then the Gabor filter is set up to perform online real-time detection, and the speed and the accuracy of the online detection can be ensured. A plurality of defects such as block defects and linear defects can be detected accurately and efficiently by using odd symmetry and even symmetry Gabor filters separately. In the circumstance of performing high-speed and real-time image acquisition by using a linear array camera, the detecting speed can be improved efficiently and the undetected rate and false detecting rate can be lowered.

Description

A kind of cloth surface flaw detection method based on wavelet neural network

Technical field

The present invention relates to a kind of Fabric Defect real-time vision detection method based on machine vision, specifically refer to that one is under array light source, cloth surface flaw industry spot high speed being transmitted by line-scan digital camera detects the also image detecting method of instant recording.

Background technology

In industrial processes, along with improving constantly of technical merit, market also promotes again and again to the requirement of product quality.In textile industry, the quality testing of cloth requires along with this development trend is further strict.But because textile output continues to increase, production line industrialization level promotes, traditional manual detection method has not caught up with the speed of automation development, be limited by supervisory personnel's subjective factor and the state of mind, and exist the inferior positions such as desk checking speed is slow, cost is high, standardization level is low, false drop rate is large, detecting quickly and accurately textile flaw becomes problem demanding prompt solution in production run.

In the face of such demand, some external large enterprises are in the industrial application that has had certain scale, main representative products has the IQ-TEX4 automatic on-line detecting system of EVS company of Israel, the Cyclops automatic on-line fabric detection system of BMS company of the U.S. etc., but with high costs, maintenance is difficult for, and does not generally promote and is suitable at home.At present, researcher mainly adopts based on methods such as statistical method, frequency domain converter technique, modellings cloth image is processed, in the hope of flaw accurately being detected, because cloth surface is with interference of texture, flaw kind is complicated, and correctly extracting defect areas becomes the Focal point and difficult point in cloth surface detection.

Due in testing process, go out cloth speed fast, cloth breadth is larger, accuracy of detection requires high, select high resolving power and be applicable to high speed acquisition process line-scan digital camera more and more become the detection mode of main flow as image acquisition sensor.

Summary of the invention

The object of the invention is to propose a kind of Fabric Defect detection method based on machine vision of highly versatile, replaces the manual detection method of traditional inefficiency.

For this object, the present invention is achieved through the following technical solutions:

Off-line state:

(1) utilize the flawless cloth image of line-scan digital camera Real-time Obtaining, the parameters such as transfer rate, collected by camera frequency and the camera aperture focal length of adjusting cloth, obtain flawless cloth image sequence in real time as sample.

(2) sample image obtaining is carried out to medium filtering processing to suppress noise remove noise spot, utilize the contrast of histogram equalization enhancing image to highlight texture.

(3) build three layers of Architecture of Feed-forward Neural Network, adopt the excitation function of an imaginary part Gabor small echo as hidden layer, build parameter vector group.

(4) utilize Levenberg-Marquardt (LM algorithm) to solve optimized parameter for each Gabor small echo, finally obtain mutually corresponding odd symmetry Gabor bank of filters and even symmetry Gabor bank of filters.

Presence:

(1) camera parameter of maintenance off-line state, Real-time Obtaining cloth image to be measured.Process and remove noise point by medium filtering, utilize histogram equalization to highlight the texture of cloth image, in each sampling period, gather cloth image formation image sequence and detect.

(2) the odd symmetry Gabor bank of filters obtaining during respectively with off-line training and even symmetry Gabor bank of filters, to picture filtering processing to be detected, detect block and wire flaw.

(3) the filtering result obtaining is carried out to fusion treatment, and to fused images smothing filtering and binaryzation, finally obtain defect areas.

The invention has the beneficial effects as follows: the invention provides a kind of image flaw detection method based on Gabor wave filter and wavelet neural network, texture information for cloth image can effectively extract, and utilize off-line training to shorten the execution time of algorithm, the odd symmetry Gabor wave filter building can well detect block flaw, even symmetry Gabor wave filter has very large advantage in the time detecting linear edge, in the face of different classes of flaw can ensure to be detected as power.Under the condition with line-scan digital camera high speed and real time sampling image, can effectively improve detection speed, reduce undetected and false drop rate.

Brief description of the drawings

Fig. 1 total system design of graphics of the present invention

Fig. 2 algorithm overall flow of the present invention figure

Embodiment

For making the object, technical solutions and advantages of the present invention etc. clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in more detail.

Basic object of the present invention is the surface blemish of inspection cloth, is divided into off-line training process and online testing process, and as shown in Figure 1, algorithm overall flow as shown in Figure 2 for the constructed in hardware of device.In off-line training process, by indefectible sample image is carried out to wavelet network algorithm process, and utilize the optimizing of LM algorithm iteration to obtain optimized parameter group, construct corresponding odd symmetry Gabor bank of filters and even symmetry Gabor bank of filters.In online testing process, utilize the bank of filters obtaining to treat detected image and carry out filtering processing, then image co-registration, defect areas after the disposal of gentle filter, finally obtained.

Further, off-line training process specific implementation step is:

(1) according to desired perching precision and speed, set cloth transfer rate, camera sample frequency, adjusts camera position etc., gathers a series of images sequence as sample image.

(2.1) utilize medium filtering to do denoising to the sample image obtaining.

(2.2) image after filtering is done to histogram equalization processing.

Gray level r in piece image _kthe probability occurring is approximately:

$P_r\left(r_k\right)=\frac{n_k}{n}$ k＝0，1，2，...，L-1

Wherein n be image pixel and, n _kthat in image, gray level is r _knumber of pixels, L is gray level sum possible in image.Have the transforming function transformation function of output gray level sk to be:

$s_k=T\left(r_k\right)=\underset{j=0}{\overset{k}{\mathrm{\Σ}}}{P}_r\left(r_j\right)=\underset{j=0}{\overset{k}{\mathrm{\Σ}}}\frac{n_j}{n}$ k＝0，1，2，...，L-1

Can be r by gray level in input picture by this transforming function transformation function _keach pixel to be mapped to gray level in output image be s _krespective pixel.

(3) build wavelet neural network.

Build three layers of Architecture of Feed-forward Neural Network.Making a secondary gray level image is f (x, y), the location label that wherein (x, y) is pixel, and f is corresponding pixel value, using Gabor wavelet network to approach this image has following representation: $\hat{f}(x,y)={\mathrm{\Σ}}_{i=1}^N{w}_i{g}_0^i(x,y)+\stackrel{\‾}{f}.$

Wherein, wi is that i hidden node is to the weight between output node

$\begin{array}{c}{g}_0^i(x,y)=\exp (-\frac{1}{2}\{{\left[\frac{(x-t_x^i)\cos {\mathrm{\θ}}^i-(y-t_y^i\sin {\mathrm{\θ}}^i)}{{\mathrm{\σ}}_x^i}\right]}^2+{\left[\frac{(x-t_x^i)\sin {\mathrm{\θ}}^i-(y-t_y^i)\cos {\mathrm{\θ}}^i}{{\mathrm{\σ}}_y^i}\right]}^2\left\}\right)\\ \×\sin \left(2\mathrm{\π}{\mathrm{\ω}}_x^i\right[(x-t_x^i)\cos {\mathrm{\θ}}^i-(y-t_y^i)\sin {\mathrm{\θ}}^i\left]\right)\end{array}$

with for the translation parameters in coordinate axis, with be the radial frequency bandwidth of i hidden node Gabor small echo, θ ⁱbe the anglec of rotation of i Gabor small echo, centered by frequency.

These parameters form parameter vector $\mathrm{\η}=(t_x^i,t_y^i,{\mathrm{\θ}}^i,{\mathrm{\σ}}_x^i,{\mathrm{\σ}}_y^i,{\mathrm{\ω}}_x^i,w_i).$

(4) utilize LM algorithm to obtain optimized parameter group, obtain Optimal Gabor Filters group.

(4.1) the parameter vector group obtaining is carried out to optimizing, the training process of whole network can be expressed as optimizing makes the minimized process of energy function exactly.

Because Gabor basis function is nonopiate, need to adopt LM algorithm to carry out iteration training and obtain optimized parameter.

Step is as follows:

N=1, uses LM algorithm to find optimal value ${\mathrm{\η}}_1=(t_x^1,t_y^1,{\mathrm{\θ}}^1,{\mathrm{\σ}}_x^1,{\mathrm{\σ}}_y^1,{\mathrm{\ω}}_x^1,w_1)$

N=2, under constant prerequisite, find optimal value with LM algorithm ${\mathrm{\η}}_2=(t_x^2,t_y^2,{\mathrm{\θ}}^2,{\mathrm{\σ}}_x^2,{\mathrm{\σ}}_y^2,{\mathrm{\ω}}_x^2,w_2)$

I, N=i, ensureing η ₁, η ₂..., η _i-1under constant prerequisite, find optimal value with LM algorithm

${\mathrm{\η}}_i=(t_x^i,t_y^i,{\mathrm{\θ}}^i,{\mathrm{\σ}}_x^i,{\mathrm{\σ}}_y^i,{\mathrm{\ω}}_x^i,w_i)$

(4.2) such Gabor filter form that just obtains an even symmetry:

$g_{\mathrm{even}}(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp \{-\frac{1}{2}[{\left(\frac{x}{{\mathrm{\σ}}_x}\right)}^2+{\left(\frac{x}{{\mathrm{\σ}}_y}\right)}^2\left]\right\}\×\cos \left(2\mathrm{\π}{\mathrm{\ω}}_{x}x\right)$

Odd symmetric Gabor filter form:

$g_{\mathrm{odd}}(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp \{-\frac{1}{2}[{\left(\frac{x}{{\mathrm{\σ}}_x}\right)}^2+{\left(\frac{x}{{\mathrm{\σ}}_y}\right)}^2\left]\right\}\×\sin \left(2\mathrm{\π}{\mathrm{\ω}}_{x}x\right)$

Online testing process specific implementation step is:

(1) camera parameter of maintenance off-line state, Real-time Obtaining cloth image to be measured, gathers cloth image formation image sequence and detects in each sampling period

(2) process and remove noise point by medium filtering, utilize histogram equalization to highlight the texture of cloth image.

(3) to testing image process Gabor bank of filters filtering processing, by the each image co-registration obtaining, obtain defect areas.

(3.1) Gabor bank of filters testing image and off-line step being obtained is carried out filtering processing:

The Gabor filter form of even symmetry:

$g_{\mathrm{even}}(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp \{-\frac{1}{2}[{\left(\frac{x}{{\mathrm{\σ}}_x}\right)}^2+{\left(\frac{x}{{\mathrm{\σ}}_y}\right)}^2\left]\right\}\×\cos \left(2\mathrm{\π}{\mathrm{\ω}}_{x}x\right)$

Odd symmetric Gabor filter form:

$g_{\mathrm{odd}}(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp \{-\frac{1}{2}[{\left(\frac{x}{{\mathrm{\σ}}_x}\right)}^2+{\left(\frac{x}{{\mathrm{\σ}}_y}\right)}^2\left]\right\}\×\sin \left(2\mathrm{\π}{\mathrm{\ω}}_{x}x\right)$

(3.2) be detected as power in order to improve, reduce erroneous judgement, need to merge the filtering result of two Gabor wave filters, with regard to Defect Detection, need to weaken background texture and strengthen the response of defect areas, first by the Output rusults (O of two wave filters _even(x, y), O _odd(x, y)) difference normalized:

$O^{\′}(x,y)=\frac{O(x,y)-\min \left(O(x,y)\right)}{\max \left(O(x,y)\right)-\min \left(O(x,y)\right)}$

(3.3) image is merged:

F(x，y)＝O′ _even(x，y)+O′ _odd(x，y)-O′ _even(x，y)×O′ _odd(x，y)

To the image denoising processing obtaining, finally separate the bianry image that obtains highlighting flaw.

(4) if this two field picture no one be defect areas, do not preserve this image sequence.Continue to detect next frame image, if flaw appears in this two field picture, preserve this flaw image and positional information to structure and continue to detect lower piece image.

Claims (4)

1. the online visible detection method of a Fabric Defect, it is characterized in that: by the combination of Gabor wave filter and wavelet neural network, effectively extract the Width information of cloth surface texture, for same kind cloth, training is asked for and is built Gabor wave filter after optimum solution and carry out detecting in real time online; For different types of flaw, corresponding selection odd symmetry, even symmetry Gabor wave filter ensure that block flaw and wire flaw detect accurately and efficiently; Specifically comprise following step:

(1) structure of wavelet neural network in off-line learning process, obtains the parameter vector of cloth surface $\mathrm{\η}=(t_x^i,t_y^i,{\mathrm{\θ}}^i,{\mathrm{\σ}}_x^i,{\mathrm{\σ}}_y^i,{\mathrm{\ω}}_x^i,w_i)$ Information;

(2) in off-line learning process, utilize LM algorithm iteration to ask for optimized parameter group, build Optimal Gabor Filters group;

(3) in online testing process, testing image is passed through to Gabor bank of filters filtering processing, by the each image co-registration obtaining, obtain defect areas.

2. a kind of online visible detection method of cloth surface flaw according to claim 1, is characterized in that: accurately the asking for of parameter vector in described step (1), comprises the following steps:

Build three layers of Architecture of Feed-forward Neural Network, establishing a width gray level image is f (x, y), the location label that wherein (x, y) is pixel, and f is corresponding pixel value, using Gabor wavelet network to approach this image has following representation: $\hat{f}(x,y)={\mathrm{\Σ}}_{i=1}^N{w}_i{g}_0^i(x,y)+\stackrel{\‾}{f};$

Wherein, w _ibe i hidden node to the weight between output node:

$\begin{array}{c}{g}_0^i(x,y)=\exp (-\frac{1}{2}\{{\left[\frac{(x-t_x^i)\cos {\mathrm{\θ}}^i-(y-t_y^i\sin {\mathrm{\θ}}^i)}{{\mathrm{\σ}}_x^i}\right]}^2+{\left[\frac{(x-t_x^i)\sin {\mathrm{\θ}}^i-(y-t_y^i)\cos {\mathrm{\θ}}^i}{{\mathrm{\σ}}_y^i}\right]}^2\left\}\right)\×\\ \sin \left(2\mathrm{\π}{\mathrm{\ω}}_x^i\right[(x-t_x^i)\cos {\mathrm{\θ}}^i-(y-t_y^i)\sin {\mathrm{\θ}}^i\left]\right),\end{array}$

with for the translation parameters in coordinate axis, with be the radial frequency bandwidth of i hidden node Gabor small echo, θ ⁱbe the anglec of rotation of i Gabor small echo, centered by frequency;

These parameters form parameter vector $\mathrm{\η}=(t_x^i,t_y^i,{\mathrm{\θ}}^i,{\mathrm{\σ}}_x^i,{\mathrm{\σ}}_y^i,{\mathrm{\ω}}_x^i,w_i).$

3. a kind of online visible detection method of cloth surface flaw according to claim 1, is characterized in that:

The structure of Optimal Gabor Filters group in described step (2), comprises the following steps:

The first step, the parameter vector group obtaining is carried out to optimizing, the training process of whole network can be expressed as optimizing makes the minimized process of energy function exactly;

Because Gabor basis function is nonopiate, need to adopt LM algorithm to carry out iteration training and obtain optimized parameter;

Step is as follows:

1, N=1, uses LM algorithm to find optimal value ${\mathrm{\η}}_1=(t_x^1,t_y^1,{\mathrm{\θ}}^1,{\mathrm{\σ}}_x^1,{\mathrm{\σ}}_y^1,{\mathrm{\ω}}_x^1,w_1)$

2, N=2, under constant prerequisite, find optimal value with LM algorithm ${\mathrm{\η}}_2=(t_x^2,t_y^2,{\mathrm{\θ}}^2,{\mathrm{\σ}}_x^2,{\mathrm{\σ}}_y^2,{\mathrm{\ω}}_x^2,w_2)$

……

I, N=i, ensureing η ₁, η ₂..., η _i-1under constant prerequisite, find optimal value with LM algorithm

${\mathrm{\η}}_i=(t_x^i,t_y^i,{\mathrm{\θ}}^i,{\mathrm{\σ}}_x^i,{\mathrm{\σ}}_y^i,{\mathrm{\ω}}_x^i,w_i);$

Second step, build and obtain according to the optimized parameter group that obtains:

The Gabor filter form of even symmetry:

$g_{\mathrm{even}}(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp \{-\frac{1}{2}[{\left(\frac{x}{{\mathrm{\σ}}_x}\right)}^2+{\left(\frac{x}{{\mathrm{\σ}}_y}\right)}^2\left]\right\}\×\cos \left(2\mathrm{\π}{\mathrm{\ω}}_{x}x\right)$

Odd symmetric Gabor filter form:

$g_{\mathrm{odd}}(x,y)=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\exp \{-\frac{1}{2}[{\left(\frac{x}{{\mathrm{\σ}}_x}\right)}^2+{\left(\frac{x}{{\mathrm{\σ}}_y}\right)}^2\left]\right\}\×\sin \left(2\mathrm{\π}{\mathrm{\ω}}_{x}x\right)$

As the optimized parameter group of presence.

4. a kind of online visible detection method of cloth surface flaw according to claim 1, is characterized in that: accurately the obtaining of defect areas in described step (3), comprises the following steps:

First by the Output rusults (O of two wave filters _even(x, y), O _odd(x, y)) difference normalized:

$O^{\′}(x,y)=\frac{O(x,y)-\min \left(O(x,y)\right)}{\max \left(O(x,y)\right)-\min \left(O(x,y)\right)}$

Merge again:

F(x，y)＝O′ _even(x，y)+O′ _odd(x，y)-O′ _even(x，y)×O′ _odd(x，y)

To the image denoising processing obtaining, finally separate the bianry image that obtains highlighting flaw.