A Robust Chinese License Plate Detection and Recognition Systemin Natural Scenes

Abstract

In natural scenes, the license plates (LP) obtained typically have different angles and character lengths, while most LPDR systems are based on license plate front view and fixed length license plates. For the above problems of Chinese license plates, in this paper, we proposed a robust LPDR system, which detects and rectifies oblique license plates in the detect stage. In LP recognition stage, we designed a pre-trained recognition network based on CNN and RNN, which improves the accuracy of Chinese character recognition in license plates, while avoiding character segmentation and license plate length limitation. Our experiment results show that our LPDR system performs well in Chinese license plates dataset CCPD, especially CCPD-Rotate and CCPD-Tilt sub-dataset, with recognition rate of 92.7% and 93.5%, respectively.