A weighted multi‐source domain adaptation approach for surface defect detection

Abstract

This paper proposes a weighted multi-source domain adaptation (MSDA) method for industrial surface defect detection. The domain adaptation method is usually used to solve inaccurate results due to lacking target training samples. But in the scene of surface defects detection, this method is not adequate because the inspection samples often contain complex texture features. To get better performance, in this paper, we extend the single-source domain adaptation detection method to the multi-source domain. At the same time, we weighted different source domains samples during the adaptive training process, and prioritize the alignment of the target domain with the most similar source domain. The experimental results show that our proposed method performs well on the target dataset which improves existing methods' limitations for detecting surface defects with complex textures.