Research On The Key Technologies Of Fabric Defect Recognition

Fabric quality control plays a very crucial role in the textile industry, and fabricdefect is an important factor affecting the quality of fabric. Currently, fabric defectrecognition is mainly by the traditional human offline inspection, which is in heavylabor intensive, slow detection speed and low detection accuracy. As the rapiddevelopment of computer technology and pattern recognition technology, fabric defectautomatic recognition is an inevitable trend in production quality control of the textileindustry. At present, research of fabric defect automatic recognition has made someachievements. But as the problems of image acquisition affect by light change and noise,and the defects have more categories, the fabric defect recognition is still a challengingresearch topic. In this thesis, we mainly investigate the defect detection andclassification algorithms. We introduce the front theories of pattern recognition in recentyears into fabric defect recognition, and further study the recognitions for10categoriescommon defects. The main contributions of this dissertation are as follows.1. According to the problems that the number of fabric defect category is large anda unity method is only effective for some types of defects, a novel fabric defectdetection algorithm is studied which combines wavelet multi-scale product andmathematical morphology. First, defect image is decomposed into sub-images using thenonsubsampled wavelet transform. Then, defect shape features are obtained bymathematical morphology operations over the low frequency sub-image. Waveletmulti-scale products methods are adopted to suppress the noise and enhance defect edgelinear features through the high frequency sub-images. Finally, the weighted averagefusion algorithm is used to get the result. The experiments are from both subjective andobjective evaluation. Compared with the classical Gabor and wavelet transformalgorithms, the proposed algorithm is fast and effective for fabric defect detection, andthe comprehensive performance of the algorithm is superior to contrast algorithms.2. Considering the advantage of decomposition coefficients in NonsubsampledContourlet transform (NSCT) of fabric images better describing the contourcharacteristics, two novel algorithms for detection of fabric defect images based onNSCT, are presented.(1) A fabric defect detection algorithm based on the standarddeviation of NSCT subbands is proposed. The optimal sub-band of NSCT is obtained bythe variance cost function. As difference of the coefficients’ between defect andnon-defective regions in the sub-band is smaller, the segmentation threshold is difficult to obtain. The standard deviation method can effectively solve this problem to getaccurate results. The algorithm is more accurate positioning of the defect and is with thesmall amount of calculation.(2) A fabric defect detection approach is presented with theGaussian Mixture Model (GMM) based on NSCT. The optimal sub-band of NSCT isacquired by the cost function method. Then, parameters of defect and non-defectiveimages are timely estimated separately by GMM, which availably avoids evaluatingeach defect. Finally, segmentation of the defect is obtained by the maximum posteriorprobability. The algorithm does not require prior knowledge of defect images, and hasgreatly improved the performance comparing with other algorithms.3. To solve the problems that the kinds of fabric defects are many and similar kindsamong are difficult to distinguish, a fabric defect classification algorithm based on theimproved GMM parameters estimation method is proposed. The global statisticalfeature is extracted from the transform sub-bands, and the experiment shows that globalfeature is stability within the class. For the advantages of Local Binary Pattern (LBP)and Gray Level Co-occurrence Matrix (GLCM), an integration feature extractionmethod of the two methods is adopted. The feature reflects the minutiae feature of thedefect and is in lower feature dimension. Later, the minimum misclassification functionis introduced to joint estimation the GMM parameters. Finally, classification is realizedby Bayesian classifier. The experimental results indicate that the hybrid features can bebetter described characteristics of the defects. Compared with the traditionalclassification methods, this algorithm obtains the higher classification accuracy, and isless impacted on the performance as the change of the sample number.4. Sparse representation based on the over-complete dictionary is a further signalrepresentation theory. As the over-complete dictionary can be effectively captured thecharacteristics of the images, and the existing defect classification algorithms rely onthe defect segmentation result, a fabric defect recognition algorithm via dictionarylearning for sparse representation is proposed. The over-complete dictionary is obtainedby the discriminative dictionary learning method, which can better capture the defectimage’s robust feature. The imaginary Gabor function is used to get the coarse positionand the rough classification of the defect, which avoids the block operations on thewhole image. Then, sparse representation based on the over-complete dictionary isobtained through the pursuit algorithm, and the classification is realized by a linearclassifier. Finally, to count and analyze the sub-block class classification information,the recognition result is obtained. The experimental results show that the proposed algorithm not only improves the classification performance but also has the errorcorrection capability for the misclassification blocks. The sparse representation methodhas improved the stability and accuracy for defects recognition.