A method for defect detection of rail transit catenary insulators based on enhanced YOLOv10

Abstract

The reliability of rail transit catenary insulators is crucial for the operational safety of electrified railways. Automated defect detection is vital for ensuring this reliability, yet existing methods often struggle to balance high accuracy and computational efficiency, particularly for deployment on resource-constrained edge devices. For instance, while advanced models like YOLOv10 offer potential, their lightweight variants often fail to accurately detect small or irregularly shaped defects, whereas high-accuracy versions are computationally prohibitive for edge use. To address this accuracy efficiency trade-off, we propose an enhanced YOLOv10 detection architecture. Specifically, we enhance the backbone with a lightweight Efficient ViT module to effectively capture diverse insulator features and global context. We employ the Wise-IoU loss function to optimize bounding box regression for improved localization of complex defects. Furthermore, knowledge distillation (using YOLOv10-s as the teacher) is applied to boost the performance of the lightweight student model. We validated our method on a large-scale, non-public, real-world dataset collected by the railway 6C inspection system. Experimental results demonstrate that compared to the YOLOv10-n baseline, our model achieves a 2.6% increase in mAP@0.5 with only a 17.9% rise in FLOPs. Crucially, compared to the higher-performing YOLOv10-s, our model achieves comparable accuracy (only 0.5% lower mAP@0.5) while reducing FLOPs by 60.1% and parameters by nearly half. This demonstrates that our approach effectively mitigates the accuracy-efficiency trade-off, offering a practical and efficient solution for real-time edge detection of catenary insulator defects.