Lightweight approach to transformers in license plate detection

Abstract

This paper explores the adaptation of lightweight transformer architectures to the field of license plate detection, facing the critical challenge of computation efficiency versus detection accuracy trade-off. Although transformer-based models such as DETR have attained state-of-the-art performance in object detection, their high computational requirements hinder their implementation on resource-limited edge devices that are commonly utilized in real-world traffic monitoring systems. Through extensive experiments of lightweight DETR models, and in particular LW-DETR, this work demonstrates that transformer models can be appropriately optimized to conduct license plate detection with real-time efficiency. The experiments on (CCPD) Chinese City Parking Dataset validate that pre-training is effective in boosting detection performance, where models pre-trained on Object365 achieve up to 9.5 percentage point improvements in mean Average Precision (mAP). The Small model of LW-DETR achieves 67.3% mAP at 333 FPS speed, making it deployment capable on edge devices. This research contributes to the field in several ways: (1) quantifying the impact of pre-training schemes on expert detection tasks and (2) establishing clear efficiency-performance profiles for making deployment decisions. The findings offer real-world solutions to automatic license plate detection for transportation infrastructure, traffic management, and security applications.