SLPNet: Towards End-to-End Car License Plate Detection and Recognition Using Lightweight CNN

Abstract

As a core component, Automatic License Plate Recognition (ALPR) plays an important role in modern Intelligent Transportation System (ITS). Due to the complexity in real world, many existing license plate detection and recognition approaches are not robust and efficient enough for practical applications, therefore ALPR still a challenging task both for engineers and researchers. In this paper, a Convolutional Neural Network (CNN) based lightweight segmentation-free ALPR framework, namely SLPNet is established, which succinctly takes license plate detection and recognition as two associated parts and is trained end-to-end. The framework not only accelerates the processing speed, but also achieves a better match between the two tasks. Other contributions includes an anchor-free LP localization network based on corners using a novel MG loss is proposed and a multi-resolution input image strategy is adopted for different tasks to balance the operation speed and accuracy. Experimental results on CCPD data set show the effectiveness and efficiency of our proposed approach. The resulting best model can achieve a recognition accuracy of 98.6% with only 3.4M parameters, while the inference speed is about 25 FPS on a NVIDIA Titan V GPU. Code is available at https://github.com/JackEasson/SLPNet_pytorch.