ScE-YOLO: an efficient approach for weld seam detection on workpiece surfaces

Abstract

Abstract With the rapid development of industrial automation, the research and application of automated weld seam grinding equipment have been receiving increasing attention. These devices not only improve production efficiency but also help ensure consistency in product quality. However, due to adverse factors such as on-site lighting and dust, accurate weld seam localization remains a key challenge for automated grinding processes. In this paper, a specialized weld seam detection network is proposed for industrial environments. A lightweight feature extraction module, G-ELAN, is employed in the backbone to reduce network computing cost while maintaining feature extraction capabilities. Then, A Spatial-Channel Feature Attention Module (SCFAM) is designed to adaptively suppress background interference and enhance detection performance. Experiments on WELD-DET dataset illustrate that our ScE-YOLO achieves the mAP of 81.8%, exceeding other compared models and surpassing the baseline YOLOv8s by 2.1%. It indicates that our network significantly enhances detection performance for weld seam detection in industrial environments. Further experiments on public NEU-DET dataset show an AP50 of 81.3%, surpassing the compared models and demonstrating its generalization capability in similar contexts.