Fabric Defect Detection for Apparel Industry: A Nonlocal Sparse Representation Approach.

With the increasing customer demand on fabric variety in fashion markets, fabric texture becomes much more diverse, which brings great challenges to accurate fabric defect detection. In this paper, a fabric inspection model, consisting of image preprocessing, image restoration, and thresholding operation, is developed to address the woven fabric defect detection problem in the apparel industry, especially for fabric with complex texture and tiny defects. The image preprocessing first improves the image contrast in order to make the details of defects more salient. Based on the learned sub-dictionaries, a non-locally centralized sparse representation model is adopted to estimate the non-defective version of the input images, so that the possible defects can be easily segmented from the residual images of the estimated images and the inputs by thresholding operation. The performance of the proposed defect detection model was evaluated through extensive experiments with various types of real fabric samples. The proposed detection model was proved to be effective and robust, and superior to some representative detection models in terms of the detection accuracy and false alarms.