YOLOv10 and RT-DETR for Weld Defect Detection: Experimental Evaluation

Abstract

Automated weld defect inspection is essential for improving product quality, reducing inspection costs, and minimizing human error in industrial manufacturing. Although deep learning-based object detectors have demonstrated promising performance, comprehensive comparisons under identical experimental conditions remain limited. This study presents a systematic evaluation of YOLOv10 and RT-DETR architectures for weld defect detection using the publicly available Weld Defects (WD) dataset. Five object detection models, namely YOLOv10-N, YOLOv10-L, YOLOv10-X, RT-DETR-L, and RT-DETR-X, were trained and evaluated under identical hyperparameter settings. In addition, the influence of different training–validation split ratios on YOLOv10 performance was investigated. Detection performance was assessed using Precision, Recall, F1-score, mAP@50, and mAP@50–95, while qualitative analyses were conducted through representative detection results and class activation mapping (CAM) visualizations. Experimental results showed that RT-DETR-X achieved the highest localization performance with a mAP@50–95 of 0.458 and the best F1-score, whereas YOLOv10-X obtained the highest precision (0.768) while maintaining competitive overall performance. Explainability analysis further demonstrated that RT-DETR-X generated more spatially coherent attention maps, whereas YOLOv10-X produced more localized but occasionally fragmented activations. The findings provide a comprehensive comparison of CNN- and Transformer-based detectors and offer practical guidance for selecting suitable deep learning models for reliable automated weld defect inspection in industrial applications.