Detection of varied defects in diverse fabric images via modified RPCA with noise term and defect prior

The purpose of this paper is to present a novel method for fabric defect detection.

The method based on joint low-rank and spare matrix recovery, since patterned fabric is manufactured by a set of predefined symmetry rules, and it can be seen as the superposition of sparse defective regions and low-rank defect-free regions. A robust principal component analysis model with a noise term is designed to handle fabric images with diverse patterns robustly. The authors also estimate a defect prior and use it to guide the matrix recovery process for accurate extraction of various fabric defects.

Experiments on plain and twill, dot-, box- and star-patterned fabric images with various defects demonstrate that the method is more efficient and robust than previous methods.

The authors present a RPCA-based model for fabric defects detection, and show how to incorporate defect prior to improve the detection results. The authors also show that more robust detection and less running time can be obtained by introducing a noise term into the model.