Hot Rolled Strip Deviation Detection Based on Improved YOLOv5

Strip deviation is a major challenge in hot rolling production. It can cause significant economic losses and safety hazards. Accurate measurement of the deviation is further complicated by the rapid motion of the strip. To solve the problem, a YOLOv5 model is proposed in this study. The model combines the C3 module with the FasterNet architecture and the EMA mechanism to improve the recognition performance. In addition, a PFCA feature extraction module is proposed to enhance the extraction of feature information. Finally, a new loss function is adopted. The original model CIoU loss function is replaced by SIoU, and Inner-IoU is integrated into a new Inner-SIoU loss function. This adjustment facilitates network convergence and improves detection accuracy. Comparative experiments show that the model's mAP improves by 2.7% over the baseline network. The detection speed can reach 30 FPS to meet the real-time requirements.