基于卷积神经网络的车牌号码识别方法  

License plate number recognition method based on convolution neural network

作　　者：王世芳 李玉龙 WANG Shi-fang;LI Yu-long(School of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,Anhui,China;School of Information Science and Engineering,Ningbo University,Ningbo 315000,Zhejiang,China)

机构地区：[1]安徽工程大学电气工程学院,安徽芜湖241000 [2]宁波大学信息科学与工程学院,浙江宁波315000

出　　处：《长安大学学报（自然科学版）》2023年第4期106-117,共12页Journal of Chang’an University(Natural Science Edition)

基　　金：高节能电机及控制技术国家地方联合工程实验室基金项目(KFKT201510);安徽工程大学校青年基金项目(2015YQ15)。

摘　　要：在自然道路交通环境中,车牌定位检测与识别是实现智慧交通系统的关键技术之一。为了解决传统方法受到分割字符失败以及暗光、大角度倾斜等不利因素影响导致车牌号码误判率较高而识别率较低的问题,提出一种使用单阶段目标检测算法(you only look once v5,YOLOv5)结合基于深度神经网络(license plate recognition neural networks,LPRNet)的车牌识别方法,使用中国停车数据集(Chinese city parking dataset,CCPD)以及自建的数据集进行训练及试验。首先YOLOv5网络利用跨阶段局部网络(cross stage partial network,CSPNet)结构进行图片特征提取,通过多尺度特征信息融合,采用完备交并比(complete intersection over union,CIOU_Loss)损失函数与非极大值抑制(non-maximum suppression,NMS)联合得到预测框,定位车牌位置。轻量级的LPRNet网络无需字符分割,使用连接时序分类(connectionist temporal classification,CTC)解决车牌字符识别问题。在YOLOv5+LPRNet网络的基础上增加改进算法,通过非线性图像增强来恢复低照度下丢失的车牌信息,利用双边滤波算法滤除背景噪声的同时保留车牌边缘信息,输入车牌矫正网络,通过透射变换得到最终车牌照片。为验证该算法的有效性,在夜间环境、角度倾斜、雨雪雾天等多种场景下开展试验。研究结果表明:车牌识别模型的平均识别准确率均超过95%,识别速度平均达到32帧/s;与Easy PR、Hyper LPR、Faster-RCNN+LPRNet和YOLOv3+LPRNet模型相比,提出方法的识别准确率和召回率均得到提高,分别达到97.65%和96.74%;提出方法在道路交通复杂的场景中表现出较强的鲁棒性,识别速度上也有明显优势。License plate localization,detection,and recognition in natural road traffic environments are key technologies for implementing intelligent transportation systems.To address the problem of high misrecognition rates and low recognition rates resulting from traditional methods that affected by factors such as failed character segmentation and adverse lighting and large angle inclination,a license plate recognition method was proposed,which combines the YOLOv5(you only look once v5)network with the LPRNet(license plate recognition neural networks)network,CCPD(Chinese city parking dataset)and a self-built dataset were used for training and experiments.First,the YOLOv5 network utilized the CSPNet(cross stage partial network)structure for image feature extraction,and through the fusion of multi-scale feature information,the prediction box was obtained through the CIOU_Loss(complete intersection over union)loss function and the NMS(non-maximum suppression)to locate the position of the license plate.The lightweight LPRNet network did not require character segmentation,but instead used CTC(connectionist temporal classification)to solve the problem of license plate character recognition.To validate the effectiveness of our algorithm,experiments in various scenarios,such as nighttime environments,angle inclination,and rainy,snowy,and foggy weather were conducted.The results show that the average recognition accuracy of the license plate recognition model exceeded 95%,with an average recognition speed of 32 frames/s.Compared with the Easy PR,Hyper LPR,Faster-RCNN+LPRNet,and YOLOv3+LPRNet models,the proposed method improves the recognition accuracy and recall rate,reaching 97.65%and 96.74%,respectively.The improved license plate recognition method exhibites strong robustness even in complex road traffic scenarios and has a significant advantage in recognition speed.This text concludes that the proposed method achieves high accuracy and robustness in license plate recognition,and has a faster recognition speed compared to othe

关 键 词：交通工程 车牌识别 LPRNet YOLOv5 车牌定位 

分 类 号：U491.116[交通运输工程—交通运输规划与管理]