PL-YOLO: A Lightweight Rail Defect Detection Algorithm

As a critical component of railway tracks, the condition of rails directly affects the safety and stability of train operations. Existing intelligent detection algorithms have been developed for rail defect detection and achieved promising results in simple scenarios. But existing algorithms still face challenges in handling diverse defect types, complex backgrounds, and limited computational resources in practical deployment. To address these issues, this paper proposes PL-YOLO, a lightweight rail defect detection algorithm. The proposed algorithm introduces two key enhancements. First, we integrate the Pinwheel-shaped Convolution into shallow layers to improve feature extraction for diverse defect types. Second, we propose the LIA_C2f module to preserve detection accuracy while significantly reducing computational complexity. Experimental results on the rail defect dataset demonstrate that PL-YOLO achieves an mAP @0.5 of 94.8% and a detection speed of 83 frames per second. Compared with the contrast algorithm, PL-YOLO achieved the highest detection accuracy with the smallest number of parameters and GFLOPs. PL-YOLO demonstrates excellent real-time performance and potential for engineering applications and contributes to the intelligent development of railway maintenance systems.