Rail track surface defect detection algorithm integrating MobileNetv3 and Transformer

Abstract

Surface damage on railway rails is a critical factor affecting train safety. To address the low efficiency and poor precision in detecting small defects like cracks and rust, a rail damage detection algorithm combining MobileNetv3 and Transformer is proposed. This algorithm embeds spatial coordinate information into channel attention and integrates the CA-Benck module into MobileNetv3 to enhance feature extraction and generalization. Then, Utilizing MobileNetv3, CA-Bneck, and the Transformer encoding module, we crafted a streamlined backbone network dubbed MobileNetV3-CATr. Finally, a BiFPN-Lite module is added to capture more defect features without increasing complexity. Through the YOLO Head, it outputs rail damage information. Experiments on our rail dataset show that our model achieves a mAP of 93.8% at 19.5 FPS, outperforming YOLOv5 by 6.5%, enabling high-precision detection of rail surface damage.