+86 138 1226 9927


Home  >  News

Introduction Of License Plate Recognition LPRNet

Apr. 22, 2019

Automatic License Plate Recognition is a challenging and important task for traffic management, digital security monitoring, vehicle identification, and parking management in large cities. Due to the blurring of the picture, the poor lighting conditions and the variability of the license plate numbers, physical influences (deformation), weather conditions and other factors make the task of license plate recognition more complicated.


A powerful license plate recognition system should be able to overcome the variability of the environment and ensure the accuracy of recognition. In other words, the system should be able to be used in the natural environment.

This paper solves the problem of license plate recognition and introduces the LPRNet algorithm. This method eliminates the need for pre-cutting of characters and subsequent identification of individual characters. In this article, we did not consider the problem of license plate detection and positioning. However, for our special case, it can be done by LBP cascade.

LRPNet is based on deep convolutional neural networks. Recent studies have demonstrated the effectiveness and superiority of convolutional neural networks in the field of machine vision, including image classification, object detection, and semantic segmentation. However, running it on an embedded device remains a challenging task.

LPRNet is an extremely college-based neural network that requires only 0.34 GFLops for every single forward propagation. Moreover, our model is deployed on the Core i7-6700k SkyLake CPU for end-to-end training and recognition with high accuracy. In addition, LPRNet can be partially ported to the FPGA, which frees up the CPU's computing power on other parts of the pipeline. Our main contributions can be summarized as follows:

LPRNet is a real-time framework for high-quality license plate recognition. Its support template and variable-length independent license plate LPR adopt end-to-end identification method without pre-cutting characters.

LPRNet is the first license plate real-time identification method that does not use RNN, which can be deployed on a variety of different platforms, including embedded devices.

LPRNet's application in real-world traffic monitoring video shows that our approach is powerful enough to handle difficult problems such as viewing angles and camera distortions, poor lighting conditions, and changes in viewing angles.