Lightweight Track surface defect detection method based on improved YOLOv5

The detection of small and clustered defects in railway track surfaces remains a challenge due to the limitations in real-time accuracy of current methods. To address this, we introduce Ghost-YOLOv5-Rail, a novel algorithm based on YOLOv5, for defect recognition. A new dataset Rail-2k was developed to overcome data scarcity. The CBS and C3 modules in the backbone were replaced with GhostNet, and the ECA module was integrated into the C3 module to replace the SE module, forming the ECA-C3Ghost module to mitigate dimensionality reduction issues. Additionally, a CBAM attention mechanism was introduced in the NECK to enhance feature extraction. Comparative experiments demonstrate a 58.6% reduction in FLOPs (from 109.1 to 45.2), a 3.16% increase in mean average precision, and a detection speed of 80.6 FPS. These results highlight the potential of Ghost-YOLOv5-Rail to advance real-time track defect detection accuracy and efficiency.