<![CDATA[Vehicle license plate detection is a key component of intelligent transportation systems, with applications in automated toll collection, traffic surveillance and security systems. This task is challenging due to variations in illumination, motion blur, occlusion, diverse license plate formats, and camera viewpoints, requiring object detection models to balance accuracy and computational efficiency, especially for real-time deployment. Recent developments in deep learning have introduced different detection paradigms, particularly convolutional neural network–based and transformer-based architectures. This paper presents a systematic comparison between a transformer-based detector, RT-DETR, and a convolutional neural network–based detector, YOLO v11, for vehicle license plate detection, focusing on accuracy–efficiency trade-offs. Experimental results show that both models achieve very high detection accuracy, with an mAP@50 of 0.995. RT-DETR achieves slightly higher performance in terms of mAP@50–95 (0.912 vs. 0.907), Precision (1.000 vs. 0.999), Recall (0.992 vs. 0.984), and F1-score (0.996 vs. 0.991). However, this accuracy advantage is accompanied by higher computational cost. RT-DETR requires an average inference time of 37.99 ms per image, corresponding to an estimated throughput of approximately 25 FPS, whereas YOLO v11 achieves a significantly lower inference time of 7.7 ms, corresponding to an estimated throughput of 129.87 FPS under the same input resolution. In addition, RT-DETR requires a longer training time (1 h 52 min) compared to YOLO v11 (1 h 22 min). Overall, this study shows that RT-DETR provides slightly higher detection accuracy, while YOLO v11 is more appropriate for real-time license plate detection due to its lower computational cost.]]>
