A COMPARATIVE EVALUATION OF YOLO11-L AND RT-DETR-L FOR RAILWAY FASTENER DEFECT DETECTION

Abstract

Rail fasteners hold the running rail to its supporting structure, and their condition has a direct bearing on track geometry, ride quality, and operational safety. Manual inspection of fasteners remains slow, labor-intensive, and inconsistent, which has motivated a sustained shift toward automated, vision-based condition monitoring. Deep learning object detectors now dominate this task, yet most published studies modify a single convolutional model, and controlled comparisons between the latest CNN-based and transformer-based real-time detectors on fastener imagery remain scarce. This paper evaluates two state-of-the-art detectors, YOLO11-L and RT-DETR-L, under identical experimental conditions for multi-class railway fastener defect detection. Both models were trained with the Ultralytics framework on a five-class dataset covering good fasteners, good joints, defective joints, missing fasteners, and missing dogscrews, using a 640 × 640 input resolution, 50 training epochs, a batch size of 8, and a single NVIDIA Tesla T4 GPU. YOLO11-L achieved a mAP@50 of 93.8% against 92.3% for RT-DETR-L, together with higher recall (93.4% vs. 91.3%), marginally faster inference (16.8 ms vs. 17.2 ms per image), and approximately 31% shorter training time (1.51 h vs. 2.18 h). RT-DETR-L returned higher precision (89.2% vs. 87.5%), and the two models were effectively tied on the stricter mAP@50–95 metric (63.6% vs. 63.5%). The results indicate that YOLO11-L offers the more favorable accuracy–efficiency balance for recall-critical fastener inspection, while RT-DETR-L remains a competitive end-to-end alternative, requiring no non-maximum suppression, in settings where false alarms carry a higher cost. These findings provide practical guidance for detector selection in automated railway track inspection systems.