Improved YOLOv5-Based Method for Tiny Defect Detection on Steel Plate

In this paper, a novel detection method based YOLOv5 is proposed, which is suitable for detecting tiny defects on steel plate. First, we introduce the Bi-Level Routing Attention (BRA) module to learn global and local detail features in the im age simultaneously, which can more efficiently extract tiny defect feature information. Second, a feature enhancement module, Spatial Context Pyramid, is designed to enhance the tiny defect feature by contrasting the spatial context information obtained from feature fusion procedure. Extensive experimental results show that our method achieves comparable performance on the NEU-DET dataset.