TGA-YOLO: An enhanced detector for multi-scale industrial defect detection with dual attention mechanisms

TGA-YOLO: An enhanced detector for multi-scale industrial defect detection with dual attention mechanisms

연구 분야: Infrastructure

학회: Signal, Image and Video Processing

Industrial surface defect detection is crucial for quality control in smart manufacturing systems, yet faces substantial challenges stemming from the multi-scale nature of defects and their complex morphological variations. While numerous algorithms have been developed, no universal solution currently addresses all detection challenges effectively. To address these challenges, this paper proposes TGA-YOLO, an enhanced detector based on You Only Look Once version 8s (YOLOv8s), incorporating dual efficient feature-learning modules to enhance multi-scale detection capability. In TGA-YOLO, the Global Target Attention on Spatial Pyramid Pooling-Fast (GTA-SPPF) module is introduced into the backbone to improve background suppression and edge information processing. Furthermore, the Global and Channel Attention Layer (GCAL) is integrated into the neck network to refine feature representation for small, background-similar, and scale-varying targets. Experimental results on three public industrial defect datasets demonstrate that TGA-YOLO achieves competitive performance compared with conventional baseline models, showing measurable improvements in precision, recall, and mean Average Precision (mAP). These findings suggest that TGA-YOLO could offer practical value for industrial defect detection across diverse industrial scenarios.