ECM-YOLO: an edge-aware contextual multi-scale fusion network for weld surface defect detection

Abstract Weld surface defect detection is crucial for ensuring structural reliability and industrial production safety. However, weld surface defects exhibit small defect regions, blurred boundaries, irregular morphologies, and substantial scale variations, making detection challenging. To address these issues, this study proposes a novel model based on YOLOv11n, termed ECM-YOLO. First, a Structure–Edge Adaptive Fusion module (SEAF) is introduced to strengthen local structural patterns, subtle boundary cues, and contrast responses, improving the representation of small and low-contrast defects. Building on the enhanced local features, a Cross-Stage Partial Gated Bidirectional Scan–Attention Fusion module (C2GBSF) is developed to capture long-range spatial dependencies and morphological continuity, enabling robust recognition of irregular and discontinuous defects. Finally, an Efficient-RepGFPN Network (ERGNet) is constructed to preserve and aggregate the enhanced local and contextual representations across multiple feature scales through bidirectional semantic–localization interaction. Experimental results on Welding Defect-V1 dataset demonstrate that ECM-YOLO achieves mAP50 and mAP50-95 scores of 88.5% and 67.4%, improving over YOLOv11n by 4.1 and 6.1 percentage points, respectively, while reducing the parameter count to 2.30 M and maintaining a computational cost of 6.6 G FLOPs. These findings indicate that ECM-YOLO provides a constructive solution for weld surface defect detection in future practical industrial scenarios.