DualSPD-YOLO: A Dual-Stage SPD-Enhanced network with TripletAttention for High-Speed weld defect detection in Thin-Walled metal tubes

Abstract

Current defect detection systems struggle to meet the requirements for high-precision and high-speed identification of multi-scale weld defects in large-sized thin-walled metal tubes. This paper proposes an intelligent weld defect detection system for thin-walled metal tubes, centered on a novel defect detection algorithm (A Dual-Stage SPD-Enhanced Network with TripletAttention mechanism). The proposed algorithm incorporates SPDConv modules to enhance its capability in detecting small targets. The TripletAttention mechanism optimizes feature fusion and lowers computational complexity and the WIoU loss function promotes generalization ability. The proposed system achieves [email protected] of 90.2% on the thin-walled metal tube weld defects dataset according to the experimental results. Notably, it attains 73.1% accuracy rate for small weld defects which is 6.3% better than the baseline models, and a frame rate of 118 frames per second (FPS), which is a significant improvement over current methods in terms of accuracy and efficiency. This detection system can be effectively deployed for non-destructive testing of various welded structures made of thin-walled metal tube, particularly demonstrating utility in weld quality inspection and defect identification for lightweight mobility device frames, such as bicycles, motorcycles, and wheelchairs.