End-to-end efficient cascade license plate recognition system in unconstrained scenarios

Abstract

Automatic License Plate Recognition (ALPR) has become a necessity for today’s urban information and transportation systems. However, the variable and unconstrained environment in real-world applications makes the license plate (LP) detection and recognition task still challenging. In this work, we design a robust end-to-end trainable LP recognition system. It employs a simple cascade framework to achieve efficient and accurate LP detection and recognition tasks. The system corrects LPs in the LP detection phase, which greatly improves LP recognition accuracy in nonideal views. Adding the attention module (containing both channel attention and spatial attention) to the LP recognition network enhances the adaptability of the network and helps the accurate delineation of characters. Experimental results show an average recognition accuracy of 98.7% on the CCPD dataset and a very good 94.4% on the Challenge sub-dataset. In addition, simultaneous recognition of multiple LPs can be achieved on the invisible CRPD dataset, demonstrating the excellent generalization capability of our system.