This paper addresses the low efficiency of manual inspection of railway bridge diseases by proposing the YOLO-OSE algorithm. Deployable on embedded platforms and designed to work with unmanned aerial vehicle (UAV)-captured images for real-time disease detection, the algorithm optimizes the YOLOv11 network structure via online convolutional reparameterization (OREPA), SlimNeck, and efficient intersection over union (EIoU). OREPA simplifies complex modules into single convolutional operations to cut training costs and speed up training. SlimNeck, with GSConv, etc., reduces computational complexity while ensuring accuracy. EIoU enhances target localization by optimizing bounding-box regression. Trained on a UAV-captured dataset, the model boosts mAP50 for four typical diseases from 88.4% to 91.5%, with low computational and storage demands. On the Jetson AGX Orin platform, it achieves 47.3 FPS, rising to 54.1 FPS after TensorRT optimization, meeting practical inspection needs. This study enhances detection efficiency, accuracy, and safety while cutting costs, thus bolstering railway bridge operation safety. It also provides practical references for the integration of artificial intelligence and civil engineering in the field of bridge monitoring and offers technical support for the improvement of structural health monitoring efficiency through scenario-based optimization strategies. By doing so, it enriches the theoretical framework of infrastructure maintenance while delivering immediate practical solutions.
Xu et al. (Wed,) studied this question.