Triple Detector based on Feature Pyramid Network for License Plate Detection and Recognition System in Unusual Conditions

Abstract

Vehicle license plate detection and recognition is one application of computer vision which was widely deployed in traffic monitoring, road toll, and parking lot system. Its benefits are reducing service labor and increase processing accuracy. There are many studies in this area such as robotics and machine learning. With the strong development of neural networks and deep learning, application deployment becomes easier and more accurate. This paper focuses on the development of a license plate detector that supports the license plate detection and recognition system in unusual conditions under the influence of light, weather, and camera placement. The license plate detector is built based on the backbone, Feature Pyramid Network (FPN), and triple detector. The network was trained and tested on the CCPD (Chinese City Parking Dataset) dataset and achieved 96.1% of AP (Average Precision) in total, the network can outperform the state-of-the-art detection network in this field. On the other hand, the license plate detection network combined with the LPRNet network (License Plate Recognition via Deep Neural Networks) embedded by the Spatial Transformer to recognize the numbers and characters in the license plate with 98.8% of AP on CCPD-Base subset. As a result, the number plate detector exhibits superior capabilities compared to previous methodologies under unusual conditions, thereby enhancing the overall processing visibility of the system.