RA-RCNN: A Physical-Feature-Aware Adaptive Detection Network for Multi-Scale Rail Surface Defects

With the rapid expansion of high-speed railways, maintaining track structural health is vital for modern railway systems. Although deep learning has improved defect detection, models still face problems such as varying defect scales, severe background noise (e.g., lubricant residues and ferruginous oxidation), and irregular defect boundaries. To solve these problems, we introduce a new network named Rail-Adaptive-RCNN (RA-RCNN). It uses a Large Selective Kernel (LSK) backbone to dynamically adjust the Effective Receptive Field (ERF) for capturing periodic corrugation. We also added an Efficient Multi-Scale Attention (EMA) module that purifies features by suppressing noise without lowering dimensions. Finally, combining Scylla-IoU (SIoU) Loss with K-means clustering optimizes the regression of odd-shaped defects. Our experiments indicate that RA-RCNN reaches a mean Average Precision (mAP0.5) of 86.2%, outperforming the baseline Faster R-CNN by 8.8%. Corrugation detection specifically reached 91.4%. With a processing speed of 26 FPS, this method effectively meets the practical needs of real-time automated railway maintenance.