AWT ‐YOLO: Lightweight Rail Surface Defect Detection Algorithm Based on YOLOv11 Improvement

ABSTRACT

Railway safety and operational reliability critically depend on timely and accurate inspection of rail infrastructure. Nevertheless, existing visual neural network‐based inspection methods frequently fail to achieve an optimal balance between detection accuracy and real‐time performance, particularly when identifying small and intricate surface defects. To mitigate this challenge, this paper designs an improved YOLOv11 model called AWT‐YOLO for rail surface damage detection. The proposed model incorporates three pivotal improvements. First, wavelet convolution (WTConv) is employed to replace the C3k2 convolution in the neck structure, enabling more effective refinement of multi‐scale features and improving sensitivity to fine‐grained surface details. Second, the original C2PSA module is substituted with a Lightweight Coordinate Attention (LCA) module, which simplifies feature extraction while significantly reducing computational overhead. Finally, we adopt the Minimum Point Distance Intersection over Union (MPDIoU) loss to enhance bounding box regression accuracy and localization precision. The experimental results show that the improved AWT‐YOLO model achieves a mAP@50 of 0.957 in the public railway track label.v6i.yolov11 dataset, which is an improvement of 3.01% over the baseline YOLOv11 model. Furthermore, the real‐time performance metric FPS increases by 14.46% and 35.82% relative to YOLOv11 and YOLOv10 models, respectively, indicating that AWT‐YOLO maintains high detection accuracy while delivering superior computational efficiency. Additionally, ablation studies on each module confirm that this algorithm has good comprehensive performance and practical application value.