Detection of Small Surface Defects on High-strength Bolts Based on Improved YOLOv5

This study proposes a YOLOv5-based algorithm for detecting surface micro-defects on high-strength bolts to address low efficiency and accuracy in industrial inspections. Key enhancements include replacing CIoU with Focal-EIoU to resolve sample imbalance and accelerate training, integrating CBAM into the Backbone to improve feature extraction and reduce brightness interference, and adopting a BiFPN-based Neck network with adjusted detection scales for small defects. On a custom dataset, the algorithm achieves AP improvements of $\mathbf{1 1. 3 \%}$ (gouge), 10.8% (crazing), and 11.9% (scratch), with mAP@ 0.5 rising 11.4% to 87.9% and real-time performance maintained at 48.5 FPS, improving the accuracy and speed for detection of surface defects of bolts.