A comparative study of methods for defect detection in textile fabrics

Abstract

Fabric defect detection methods have been broadly classified into three categories; statistical methods, spectral methods and model-based methods. The performance of each method relies on the discriminative ability of texture features it uses. Each of the three categories has its own advantages and disadvantages and some researchers have recommended their combination for improved performance. In this paper, we compare the performance of three fabric defect detection methods, one from each of the three categories. The three methods are based on the grey-level co-occurrence matrices (GLCM), the undecimated discrete wavelet transform (UDWT) and the Gaussian Markov Random field models (GMRF) respectively from the statistical, spectral and model-based categories. The tests were done using the textile images from the TILDA dataset. To ensure classifier independence on the outcome of the comparison, the Euclidean distance and feed forward neural network classifiers were used for defect detection using the features obtained from each of the three methods. The results show that GLCM features allowed better defect detection than wavelet features and that wavelet features allowed better detection than GMRF features.