INTRODUCTION: The surface images of steel rails are extremely difficult to detect and recognize due to the presence of interference such as light changes and texture background clutter during the acquisition process. METHODS: To improve the accuracy of railway defects detection, a deep learning algorithm is proposed to detect the rail defects. Aiming at the problems of inconspicuous rail defects edges, small size and background texture interference, the rail region extraction, improved Retinex image enhancement, background modeling difference, and threshold segmentation are performed sequentially to obtain the segmentation map of defects. For the classification of defects, Res2Net and CBAM attention mechanism are introduced to improve the receptive field and small target position weights. The bottom-up path enhancement structure is removed from the PANet structure to reduce the parameter redundancy and enhance the feature extraction of small targets. RESULTS: The results show the average accuracy of rail defects detection reaches 92.68%, the recall rate reaches 92.33%, and the average detection time reaches an average of 0.068 s per image, which can meet the real-time of rail defects detection. DISCUSSION: Comparing the improved method with the mainstream target detection algorithms such as Faster RCNN, SSD, YOLOv3 and other algorithms, the improved YOLOv4 has excellent comprehensive performance for rail defects detection, the improved YOLOv4 model obviously better than several others in , , and F1 value, and can be well-applied to rail defect detection projects.