Rail Surface Defect Detection Based on Multi-scale Feature Fusion

To address issues such as insufficient detection accuracy and redundant model parameters in the detection of surface defects on railway tracks, a lightweight and efficient detection algorithm, EMS-YOLO, is proposed. Based on YOLOv8s, EMS-YOLO firstly improves its neck architecture with a lightweight efficient multi-scale feature fusion network EMSFPN to improve feature representation for small-scale rail surface defects. To further optimize the detection performance and reduce the number of parameters and computations, we employ Lite-DESH, a lightweight detail-enhanced shared detection head. Experimental results on the self-built Rail track surface defect dataset demonstrate that EMS-YOLO achieves an effective balance between detection accuracy and model complexity. Compared to the baseline YOLOv8s, EMS-YOLO improves mAP0.5 by 2.5%, while reducing the number of parameters by 30% and computational complexity by 35.4%. Comparative evaluations with other mainstream detection algorithms further confirm EMS-YOLO’s superior performance in complex track environments.