Research On The Key Technology Of Automated Fabric Defect Detection System

Fabric defect detection is an important procedure for quality control of modern manufacturing in thetextile industry. Fabric defect detection is usually performed by human inspectors with a low accuracy.Hence, the effective and practical automated fabric defect detection algorithm is developed using digitalimage processing technology and pattern recognition technology in this study.Plain fabric is taken as the main research object in this paper. Weaving mechanism, structuralcharacteristics, defect characteristics of fabric and texture characteristics of the fabric image are analyzedwhich are applied into the fabric defect detection algorithm as prior knowledge. In order to solve thecomplicated problem of automated fabric defect detection, three kinds of fabric defect detection algorithmare designed from different angles, regarding practicability and real-time as objective. The first is the fabricdefect detection algorithm based on projected transform for feature extraction in the spatial domain, thesecond is the fabric defect detection algorithm based on the spectral characteristic of fabric image in thefrequency domain, and the third is the fabric defect detection algorithm based on the Gabor filter usingspatial frequency analysis.In order to solve the key technology problem of automated fabric defect detection, projected transformis proposed to extract features of the fabric image making use of fabric characteristic and the method ofanomaly detection is developed to detect defects in this paper. Automated fabric defect detection scheme isdivided into two phases, which are the study phase and the detection phase. During the study phase,features of normal fabric image are extracted to get the feature data set of normal fabric and the normalrange of each feature value is acquired by statistical method. During the detection phase, the method ofanomaly detection is developed using features of fabric image to detect defect and the feature values arecomputed for a set of windows covering the image. The effect of the size of the window on the fabricdefect detection algorithm is also discussed. Analysis on the feature values of the defective fabric imageshows that each of feature values is in the normal range for normal fabric image, and one feature value atleast is abnormal for defective fabric image. Defects can be located according to the location of abnormalvalue. Testing on general fabric by this method, experimental results obtained have indicated that thescheme is suitable for the fabric image that has the legible yarns, and make the calculations simple and fastfor the algorithm to be suitable for real-time applications.The fabric defect detection algorithm based on the spectral characteristic of the fabric image isproposed in this study. The woven fabric image is operated by fast Fourier transform (FFT) and thenfiltered by the Gaussian filter designed in the frequency domain to attenuate the background texture. Afterthe inverse Fourier transform (IFT) of the output filtered image, the threshold operation is carried out tosegment defects. The power spectrum of fabric image is derived from FT function. The spectralcharacteristic of woven fabric image is analyzed in order to design Gaussian filter in the frequency domain.The periodic structures which are formed by the regular arrangement of warp and weft will result in thepeaks in the power spectrum. The first peak which is nearest to origin in the horizontal (vertical) directionhas relationship with the warp yarn (weft yarn) density of woven fabric. The physical meaning of the peaksin the power spectrum is analyzed and the relationship between peaks and fabric density is tested from bothsides of theory and experiment. The method of locating peaks accurately using fabric density is presented.The fabric defect detection algorithm based on the spectral characteristic overcomes the problem ofunstable feature extraction in the fabric defect detection algorithm based on projected transform and has theadvantage of simplicity and feasibility. The performance of the proposed algorithm has been evaluated byusing a set of woven fabric images. The experimental results have indicated that the algorithm performs very well in detecting woven fabric defects.A Gabor filters scheme is presented for unsupervised woven fabric defect detection in this paper. Thedesign for parameters of Gabor filters in the frequency domain is studied in detail. The Gabor filters aredesigned in the frequency domain by using the prior knowledge of woven fabric structure parameters,spectral characteristic of fabric image and defect characteristics in this scheme. An input woven fabricimage is filtered in the frequency domain by the Gabor filters tuned to certain frequency and orientation,which produces an output image containing the minimum amount of background texture details whileaccentuating defect details required for defect detection. A threshold can then be performed to segmentdefects from the woven fabric image. The performance of the system is evaluated on woven fabrics withdifferent types of defects. The results indicate that the scheme is available and efficient and the truedetection rate can achieve90%. The advantage of the algorithm is that the design of Gabor filterparameters is simplified using the spectral characteristic of the fabric image. The requirements of fabricdefect detection algorithm can be satisfied by only two Gabor filters, so less time is required for filtering.Gabor filter parameters are obtained according to the woven fabric structure parameters and the spectralcharacteristic of the fabric image for different fabric. The scheme is feasible and can satisfy the real timeand accurate requirement of fabric defect detection system.