Learning the Color Space for Effective Fabric Defect Detection

Fabric defect detection has been an important but challenging problem. One crucial problem lies in that the fabric colors and textures highly affect the detection performance. Especially, defects with low contrast to the fabric patterns are difficult to detect. We find that the color space is one important issue, and defects exhibit diverse discriminative levels given different color spaces. This observation inspired us to seek for an optimal color space for the defect detection problem. We propose a Color Conversion Network (CCN), which learns to project RGB colors into an optimized color space, where the fabric textures and the defective regions are the most distinguishable, to improve the defect detection performance. By introducing a contrastive loss, CCN can maximize the distance between defects and normal patterns effectively. The proposed CCN can be adopted as an add-in module that flexibly integrates with existing approaches to achieve universal improvement. Experiments demonstrate that the CCN module increases the fabric defect detection performance significantly with four state-of-the-art fabric defect detection approaches.