A Large-Angle License Plate Detection and Recognition System in Chinese Port Scenarios

Yixin Li, Saijun Wang, Zhexuan Ye, Chen Ye

Abstract: License plate detection and recognition (LPDR) holds a pivotal role in intelligent transport systems (ITS), facilitating seamless traffic surveillance and automated vehicle information retrieval. However, particularly in port environments, challenges such as spatial constraints often result in significant angles between cameras and vehicles, leading to diminished recognition efficiency. Due to the high complexity of vehicle images in port scenarios and the lack of port scenes in most public vehicle datasets, we introduce a large-scale Port Container Truck Dataset (PCTD) containing a total of 14,295 synthetic and real images. Furthermore, we present a cascading LPDR framework, leveraging the YOLOv5 detection algorithm alongside the rectification-based PP-OCRv4 model for precise license plate (LP) identification in port environments. To summarize, we develop a high-performing LPDR approach with three insights. Firstly, this system demonstrates high accuracy and robustness in complex environments, adeptly handling challenges such as blurriness, low resolution, large angles, and extreme weather conditions. Secondly, while most current LPDR systems can only recognize single-line LPs, our system is designed to effectively identify double-line LPs through a dedicated recognition module. Lastly, we propose a novel method to test LP localization accuracy and resistance to interference by creating a unique synthetic dataset of vehicles without LPs.