Abstract:

To address the issue of insufficient feature extraction and imprecise target localization in existing algorithms for detecting bridge bolt defects due to the complexity of bolt backgrounds and their small size， a bridge bolt defect recognition method based on enhanced YOLOv5s was proposed. Attention mechanisms in the backbone network was introduced to enhance the model's ability to extract bolt features and deepen its focus on global bolt characteristics. The spatial pyramid pooling structure was optimized to reduce the loss of bolt feature information. MPDIoU was employed as the bounding box regression loss function to improve the accuracy of bolt bounding box regression. The YOLO detection head was decoupled to eliminate the adverse effects of shared detection head in target detection on the regression of bounding box positions. Training and testing were conducted on 3810 self-made bolt image datasets of four typical defects： bolt rusting， bolt loosening， bolt detachment， and nut detachment， as well as normal bolts. Experimental results show that the algorithm achieves a detection accuracy of 90.8% for bolt defects， which is a 3% improvement over YOLOv5s， and a mean average precision of 92.6%， representing a 4.3% improvement over YOLOv5s. This method can be applied for intelligent recognition of bolt defects in bridges.

Key words: bridge engineering, bolt defect identification, YOLOv5s, bridge bolts