Advanced deep learning techniques for automated license plate recognition

Abstract

This research improves the capacity of automated license plate recognition (ALPR) to meet not only the needs of its methodology but also those it is confronted with in everyday situations. By combining YOLOv10 with a specially customized Tesseract OCR engine, the aim was to achieve the recognition of the Thai–Roman mixed-script car license plates, which represents a difficult and scarcely resolved problem in the literature. To ensure the system can be thoroughly tested in a wide range of scenarios, we have assembled a large-scale dataset that comprises 50,000 images and 10,000 video clips depicting different lighting and weather conditions. We have tested real-time capability on Jetson Nano and our results support the possibility of scaling for intelligent transportation systems. Comparing our experiments to the results of the latest detectors (YOLOv5, YOLOv8, YOLOv9, Faster R-CNN, SSD), we find that YOLOv10 consistently gives better results with detection accuracy of 99.16%, an F1-score of 0.992, and an inference time of 1.0 ms/image, while under severe conditions there is no significant decrease of performance. In sum, these empirical results turn the proposed system into a both novel and practical contribution for regionally adaptive ALPR research.