Implementation of Real-ESRGAN for image resolution enhancement in a YOLOv8-based vehicle license plate identification system under low-light conditions

Abstract

Vehicle license plate identification systems still face significant challenges under low-light conditions, including low image resolution, high noise levels, and decreased detection accuracy. Conventional methods such as contrast enhancement or filtering are often insufficient to recover textual details of license plates. This study implements Real-ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) for image resolution enhancement prior to license plate detection using YOLOv8. The proposed system is deployed on an NVIDIA Jetson Nano edge computing device to support real-time inference. Low-quality input images acquired under low-light conditions are first enhanced using Real-ESRGAN to restore image details and improve resolution. The enhanced images are then processed by YOLOv8 for license plate detection and character recognition. Experiments were conducted on 1,200 vehicle license plate images captured under three conditions: nighttime, underground parking, and heavy rain. Evaluation results show that Real-ESRGAN improves PSNR by 4.21 dB and SSIM by 0.12 compared with the original low-light images. License plate detection accuracy (mAP@0.5) increases from 71.32% to 94.58%, and character recognition accuracy improves from 65.47% to 91.23%. The average system response time is 1.89 s, which remains within an acceptable range for real-time smart parking applications. Overall, the proposed Real-ESRGAN–YOLOv8 framework effectively improves vehicle license plate identification under low-light conditions.