YOLO-GAFPN: An Improved YOLOV8 with Gated Attention Feature Pyramid Network for Multi-Scale Weld Defect Detection

Welding defect detection is essential for ensuring welding safety. However, traditional nondestructive testing methods often suffer from significant misdetection, particularly for multi-scale defects. In addition, existing detection methods also suffer from some shortcomings, such as degradation of small features, background noise interference, and incomplete multi-scale feature capture. Faced with the above issues, this paper proposes an improved YOLO detection model, named YOLO-GAFPN, for accurate weld defect detection. Specifically, proposed model introduces an Non-local module, which is used to capture long-distance dependencies to enhance global context awareness. Additionally, a Gated Attention Feature Pyramid Network (GAFPN) is built for enriching multi-scale feature representation, to allow the detection model to simultaneously capture global context and strengthen attention to key local information. Finally, a dynamic, non-monotonic focused loss function, named Wise-IoUv3, is introduced to adaptively adjust anchor frame weights for optimized multi-scale regression. Experiments show that proposed model achieves 73.8% [email protected] and 36.6% mAP50-95 on the industrial weld defect dataset, which represents a 3.7% and 3.4% improvement over the YOLOv8 baseline, demonstrating enhanced weld defect detection performance.