基于卷积神经网络的双行车牌分割算法  

Double-Row License Plate Segmentation with Convolutional Neural Networks

作　　者：赵汉理[1] 刘俊如 姜磊[1] 沈建冰[2] 胡明晓[1] Zhao Hanli;Liu Junru;Jiang Lei;Shen Jianbing;Hu Mingxiao(Intelligent Information Systems Institute,Wenzhou University,Wenzhou 325035;School of Computer Science&Technology,Beijing University of Technology,Beijing 100081)

机构地区：[1]温州大学智能信息系统研究所,温州325035 [2]北京理工大学计算机学院,北京100081

出　　处：《计算机辅助设计与图形学学报》2019年第8期1320-1329,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：浙江省自然科学基金(LY14F020032);浙江大学CAD&CG国家重点实验室开放课题(A1805);温州大学研究生创新基金(18-26)

摘　　要：随着智能化交通的迅速发展,自动车牌识别技术不断提高.现有大多数车牌识别技术能较好识别单行车牌字符信息,但双行车牌识别准确率较低且支持中文双行车牌的识别算法更少.为了有效地将原本仅支持单行车牌识别的算法扩展到双行车牌识别,提出一种基于卷积神经网络(CNN)的双行车牌分割算法,首先利用CNN提取车牌图像特征;然后利用特征训练多标签分类模型,将双行车牌分割为2个单行车牌.文中还构建了一个包含20多万幅中国车牌图像的数据集.基于此数据集的实验结果表明,文中算法对双行车牌自动分割准确率较高,有效地提高了双行车牌识别准确率.With the fast development of intelligent traffic,the license plate recognition technology progressively improves.Most of existing license plate recognition techniques can well recognize character information for singlerow license plates but the recognition accuracies for double-row license plates are not ideal and even less algorithms support Chinese characters.This paper introduces a doublerow license plate segmentation method with CNN,enabling efficient double-row license plate recognition for originally single-row recognition algorithms.First,this method trains a multi-label classification model with the image features extracted using CNN.Then,we use the model to automatically segment a double-row license plate into two single-row license plates.In addition,we have constructed a training and validation dataset containing more than 200 000 Chinese license plate images.The experimental results show that the proposed method has a higher accuracy in automatic segmentation of double-row license plate,thus effectively improving the accuracy of double-row license plate recognition.

关 键 词：车牌分割 车牌识别 卷积神经网络 多任务学习 

分 类 号：TP317.41[自动化与计算机技术—计算机软件与理论]