A VISION-BASED QUALITY INSPECTION SYSTEM FOR FABRIC DEFECT DETECTION AND CLASSIFICATION

Abstract

Quality inspection of textile products is an important issue for fabric manufacturers. It is desirable to produce the highest quality goods in the shortest amount of time possible. Fabric faults or defects are responsible for nearly 85% of the defects found by the garment industry. Manufacturers recover only 45 to 65% of their profits from second or off-quality goods. There is a need for reliable automated woven fabric inspection methods in the textile industry. Numerous methods have been proposed for detecting defects in textile. The methods are generally grouped into three main categories according to the techniques they use for texture feature extraction, namely statistical approaches, spectral approaches and model-based approaches. In this thesis, we study one method from each category and propose their combinations in order to get improved fabric defect detection and classification accuracy. The three chosen methods are the grey level co-occurrence matrix (GLCM) from the statistical category, the wavelet transform from the spectral category and the Markov random field (MRF) from the model-based category. We identify the most effective texture features for each of those methods and for different fabric types in order to combine them. Using GLCM, we identify the optimal number of features, the optimal quantisation level of the original image and the optimal intersample distance to use. We identify the optimal GLCM features for different types of fabrics and for three different classifiers. Using the wavelet transform, we compare the defect detection and classification performance of features derived from the undecimated discrete wavelet and those derived from the dual-tree complex wavelet transform. We identify the best features for different types of fabrics. Using the Markov random field, we study the performance for fabric defect detection and classification of features derived from different models of Gaussian Markov random fields of order from 1 through 9. For each fabric type we identify the best model order. Finally, we propose three combination schemes of the best features identified from the three methods and study their fabric detection and classification performance. They lead generally to improved performance as compared to the individual methods, but two of them need further improvement.