Abstract
Weld defect detection poses significant challenges including ambiguous boundaries, diverse defect shapes, and the requirement for precise localization. To address these issues, we propose DSF-YOLO, a novel framework specifically designed for pipeline weld defect detection. DSF-YOLO introduces three core innovations. The Dynamic Staged Fusion Feature Extraction (DSFFE) module dynamically fuses same-scale features from dual-backbone networks, progressively enhancing the representation of defect features and enabling the model to efficiently capture small-sized defects, blurred boundaries, and complex defect characteristics. The Dual Multi-Scale Feature Fusion (DMFF) module builds on the feature extraction capabilities of DSFFE and employs a dual fusion strategy to effectively aggregate global and local features, enhancing the representation of small targets and improving the separation of blurred boundaries. The decoupled head based on SENetv2-ResNeXt incorporates a multi-channel parallel processing strategy to further strengthen feature representation while inter-channel information interaction and global feature representation significantly improve classification and localization precision. Validated on an X-ray weld defect dataset containing 8 defect types, DSF-YOLO achieved an mAP50:95 of 74.7% surpassing YOLOv8-X by 1.1% and an mAP50 of 99.1% exceeding YOLOv8-X by 0.3%. Additionally, DSF-YOLO significantly reduces computational complexity, achieving a 75% reduction in FLOPs and a 47.5% decrease in parameters compared to YOLOv8-X. These results establish DSF-YOLO as an efficient and accurate solution addressing critical challenges in industrial weld defect detection with significant practical value.