Fabric Defect Detection Based on Total Variation Regularized Double Low-Rank Matrix Representation

Abstract

Fabric defect detection plays an irreplaceable role in textile quality control. Fabric images collected on industrial sites are complex and diverse, which brings great challenges to defect detection. Fabric detection algorithm based on traditional image processing method has low detection accuracy and lack of adaptability. Low rank representation model has been proved to be suitable for fabric defect detection. Normal fabric backgrounds have high redundancy and located in a low-dimensional subspace. Meanwhile, we have noticed that the defect is a region with certain edge characteristics formed by the aggregation of multiple pixels, which has high redundancy in its interior. Therefore, fabric defects can be seen as located in a low-dimensional subspace independent of the background. In this paper, a fabric defect detection algorithm based on DERF descriptors and total variation regularized double low-rank matrix representation is proposed. The characteristic matrix of the test fabric image is extracted by DERF descriptor, and the fabric image is represented as background and defect by the method of total variation regularized double low-rank representation. Experiments on two datasets show that our method has good detection performance for plain, twill and complex patterned fabrics, and is superior to other state-of-the-art method.