Research On License Plate Recognition Based On Convolutional Neural Network

As the core link of the intelligent traffic management system,the license plate recognition technology provides strong technical support and effective management methods for the realization of intelligent transportation.Although the traditional license plate detection system relies on the pictures taken by the high-definition camera to successfully recognize the license plate information,and applied to some specific scenes such as community entrances,parking Jots,etc,but for the complex license plate image,the traditional license plate recognition technology cannot accurately locate and recognize the license plate.This paper aims at complex license plate images,using deep learning algorithms and the powerful functions of computers to establish a license plate detection and recognition model based on convolutional neural network,which has important theoretical significance for the application of license plate recognition technology in the field of deep learning.The paper starts with the traditional license plate detection algorithm,by analyzing the characteristics of China’s license plates.However the license plate may be positioned incorrectly if the car body and the license plate colors and textures appear similar.In this paper,a positioning method which combines the edge detection and color information is introduced.According to the size,shape and color characteristics of the license plate,the candidate area obtained by the closed operation is screened,which can realize the positioning rapidly and accurately.This method has a good positioning effect for blue license plates in the community gates,parking lots and other occasions.However,for yellow license plates,complex image backgrounds,multiple license plates in pictures and so on,this method cannot accurately locate the license plate area.In order to improve the applicability and robustness of the license plate location algorithm,deep learning is used to detect the license plate.Regarding the plate region detection as two types of target detection problems of the license plate region and the background region,proposes a license plate region detection algorithm based on Mask R-CNN.In this paper,ResNet+FPN network model was selected on the basis of hardware to extract target features,and ResNet-50,ResNet-101,ResNet-152 three different depth ResNet networks were selected as the basic feature extraction networks for the Mask R-CNN model.Finally,the ResNet-101 network with a detection rate of 97.8%is selected as the original feature extraction network for the license plate area detection algorithm by comparing the detection rates of the three networks.Effectively solve the problems in traditional license plate detection algorithms,which improves the detection rate of license plates,network generalization ability,and algorithm robustness.A CNN model is advanced to solve the more characters category,complex background,Chinese characters and provincial abbreviation(only letters)problems which can be divided into three parts:Chinese character network,letter network and digital letter network.Then use the designed CNN model to train and recognize the license plate character dataset.And the final experimental result show that the comprehensive correct recognition rate of the three networks can reach 98.58%,which could satisfy the practical application requirements.