YOLOv8 and Faster R-CNN Performance Evaluation with Super-resolution in License Plate Recognition

Abstract

License Plate Recognition (LPR) systems are now indispensable technology for law enforcement, border control, traffic management, and parking facilities, among other sectors. They enable enhancement in security and public safety, streamline traffic management, boost staff productivity, and deliver a seamless experience for customers. However, the current model faces challenges in producing high-quality images from a fixed-angle camera to produce accurate results in recognizing characters on the license plate. Weather conditions and unfavorable LP angles can lead to low-resolution images, causing inaccuracies in recognizing the character. Therefore, improvement is needed. In the past, to solve these issues, researchers have developed Super-resolution (SR) models capable of generating high-resolution images from low-resolution counterparts. In this paper, the authors enhance the LPR technology to become an automatic solution which is called Automatic License Plate Recognition (ALPR) system by incorporating SR, aiming to automatically improve character recognition. The study comprises two phases: the detection phase and the recognition phase. In the detection phase, the authors utilize state-of-the-art object detection models, including the YOLOv8 model, and the Faster R-CNN model that uses detectron2. These models perform well. YOLOv8 achieves 93% accuracy in both train and validation datasets, and 90% in the test dataset. While Faster R-CNN achieves 71%, and 74%, respectively. In the recognition phase, the authors employ stand-alone Tesseract-OCR and SRGAN-enabled Tesseract-OCR. The end-to-end pipeline achieves a Character Error Rate (CER) of 53.9% (stand-alone Tesseract-OCR) and 51.7% (SRGAN-enabled Tesseract-OCR ). At the same time, Levenshtein distance achieves 3.6% (stand-alone Tesseract-OCR) and 3.5% (SRGAN-enabled Tesseract-OCR). This highlights the effectiveness of SRGAN in enhancing image quality and, consequently, improving the performance of OCR engines. The insights gained from this study can contribute to the development of robust license plate recognition systems for real-world deployment.