Detection Method of Concave Defect on Specular Surfaces Based on Swin Transformer

Tanaka, Kazumoto

Engineering Science Letter
Vol. 4 No. 01 (2025): Engineering Science Letter

Shallow concave defects on mirrored surfaces are difficult to detect automatically. This paper proposes a defect detection method using a deep neural network (DNN) that learns the presence or absence of distortion in the image of a stripe pattern reflected on a mirror surface. The Swin Transformer is used as the DNN to capture global features of the edges of the reflection. In the manufacturing process, the occurrence of defects is minimized, so it is difficult to collect enough defect images for training purposes. Therefore, in this paper, we show how to generate a large number of images of stripe pattern reflections using an optical simulation method. Our Swin Transformer showed high detection performance in defect detection experiments using actual mirrored parts.