In this paper, the problem of fabric defect detection using machine vision is investigated. Every inspection image is used to generate two projection signals along the horizontal and vertical planes respectively. Each of these signals is then normalized to have zero mean and unity variance. Wavelet decomposition of these normalized projection signals is used to enhance defect information. A simple thresholding operation on these wavelet coefficients extracts the defects and their localization. Experimental results presented in this paper show that this approach is highly successful in detecting variety of fabric defects and, can provide low-cost, single PC-based, solution to the web inspection problem.

published_or_final_version