一种基于GMP-LeNet网络的车牌识别方法
|
摘要
车牌识别技术是智能交通管理系统的核心,对它的研究与开发具有重要的商业前景。传统的车牌字符识别方法存在特征提取复杂的问题,而卷积神经网络作为一种高效识别算法,对处理二维车牌图像具有独特的优越性。针对传统卷积神经网络LeNet-5识别车牌图像时,存在训练数据较少、全连接层参数冗余以及网络严重过拟合等一系列的问题,设计了一种全局中间值池化(GMP-LeNet)网络,其使用卷积层代替全连接层,利用Network In Network网络中的1*1卷积核进行通道降维,全局均值池化层直接将降维后的特征图馈送到输出层。实验证明,GMP-LeNet网络能有效抑制过拟合现象,并具有较快的识别速度和较高的鲁棒性,车牌识别率达到了98.5%。
As the core of intelligent traffic management system,the research of license plate recognition technology has important business prospects.The traditional license plate character recognition method has the problem of complex feature extraction.As an efficient recognition algorithm,convolution neural network has a unique superiority in dealing with two-dimensional license plate image.When the traditional convolution neural network LeNet-5 identifies the license plate image,there is a series of problems such as less training data,redundancy of the fully connection layer and over-fitting of the network.A global intermediate pool(GMP-LeNet)network was designed,which utilizes the convolution layer instead of the fully connection layer.The 1*1 convolution kernel learning from the NIN network is used to reduce channel dimension.Then the global mean pool layer feeds the feature graph to the output layer after the dimension reducing directly.Experiments show that GMP-LeNet network can suppress the over-fitting phenomenon effectively with a faster recognition speed and the higher robustness.The final license plate recognition rate is close to 98.5%.
As the core of intelligent traffic management system,the research of license plate recognition technology has important business prospects.The traditional license plate character recognition method has the problem of complex feature extraction.As an efficient recognition algorithm,convolution neural network has a unique superiority in dealing with two-dimensional license plate image.When the traditional convolution neural network LeNet-5 identifies the license plate image,there is a series of problems such as less training data,redundancy of the fully connection layer and over-fitting of the network.A global intermediate pool(GMP-LeNet)network was designed,which utilizes the convolution layer instead of the fully connection layer.The 1*1 convolution kernel learning from the NIN network is used to reduce channel dimension.Then the global mean pool layer feeds the feature graph to the output layer after the dimension reducing directly.Experiments show that GMP-LeNet network can suppress the over-fitting phenomenon effectively with a faster recognition speed and the higher robustness.The final license plate recognition rate is close to 98.5%.
引文
[1]马爽,樊养余,雷涛,等.一种基于多特征提取的实用车牌识别方法[J].计算机应用研究,2013,30(11):3495-3499.
[2]王晓雪,苏杏丽.数字图像处理在车牌识别中的应用[J].自动化仪表,2010,31(7):22-25.
[3]陈玮,曹志广,李剑平.改进的模板匹配方法在车牌识别中的应用[J].计算机工程与设计,2013,13(5):1808-1811.
[4]KIM K B,CHO J H.Recognition System of Car License Plate using Fuzzy Neural Networks[J].Journal of the Korea Society of Computer and Information,2007,12(5):313-319.
[5]MAHDI A,FATEMEH S,TABARI Z,et al.Automatic Iranian Vehicle License Plate Recognition System Based on Support Vector Machine(SVM)Algorithms[J].Computer Engineering and Applications Journa,2013,2(1):161-174.
[6]陈扬.数字图像模式识别在车牌自动识别中的应用研究[D].天津:天津大学,2017.
[7]ABDULLAH S N H S,KHALID M,YUSOF R,et al.License Plate Recognition using Multiclusterand Multilayer Neural Networks[J].Information and Communication Tecimologies,2006,1:1818-1823.
[8]HE S,YANG C,PAN J S.The Research of Chinese License Plates Recognition Based on CNN and Length Feature[M]∥Trends in Applied Knowledge-Based Systems and Data Science.Springer International Publishing,2016:389-397.
[9]AHRANJANY S S,RAZZAZI F.A Very High Accuracy Handwritten Character Recognition System forFarsi/Arabic Digits Using Convolutional Neural Networks[C]∥Liverpool:International Conference on Bio-Inspired Computing.2010:1585-1592.
[10]赵志宏,杨绍普,马增强.基于卷积神经网络LeNet-5的车牌字符识别研究[J].系统仿真学报,2010,22(3):638-641.
[11]张立,朱玉全,陈耿.基于卷积神经网络SLeNet-5的车牌识别方法[J].信息技术,2014,11:7-11.
[12]LIU Y J,HUANG H.Car Plate Character Recognition Using a Convolutional Neural Network with Shared HiddenLayers[C]∥Chinese Automation Congress.Wuhan,2015:638-643.
[13]YUAN A Q,BAI G,JIAO L J.Offline Handwritten English Character Recognitionbased on Convolutional Neural Network[C]∥International Workshop on Document Analysis Systems.Toronto,2012:125-129.
[14]KRIZHEVSKY A,SUTSKEVER I,HINTON G E.ImageNet classification with deep convolutional neural networks[C]∥International Conference on Neural Information Processing Systems.Curran Associates Inc.2012:1097-1105.
[15]LIN M,CHEN Q,YAN S C.Network In Network[D].Banff:International Conference on Learning Representations,2014:1-10.
[16]刘万军,梁雪剑,曲海成.不同池化模型的卷积神经网络学习性能研究[J].中国图象图形学报,2016,21(9):1178-1190.
[17]郭荣艳,胡雪惠.BP神经网络在车牌字符识别中的应用研究[J].计算机仿真,2010(9):299-301.
[2]王晓雪,苏杏丽.数字图像处理在车牌识别中的应用[J].自动化仪表,2010,31(7):22-25.
[3]陈玮,曹志广,李剑平.改进的模板匹配方法在车牌识别中的应用[J].计算机工程与设计,2013,13(5):1808-1811.
[4]KIM K B,CHO J H.Recognition System of Car License Plate using Fuzzy Neural Networks[J].Journal of the Korea Society of Computer and Information,2007,12(5):313-319.
[5]MAHDI A,FATEMEH S,TABARI Z,et al.Automatic Iranian Vehicle License Plate Recognition System Based on Support Vector Machine(SVM)Algorithms[J].Computer Engineering and Applications Journa,2013,2(1):161-174.
[6]陈扬.数字图像模式识别在车牌自动识别中的应用研究[D].天津:天津大学,2017.
[7]ABDULLAH S N H S,KHALID M,YUSOF R,et al.License Plate Recognition using Multiclusterand Multilayer Neural Networks[J].Information and Communication Tecimologies,2006,1:1818-1823.
[8]HE S,YANG C,PAN J S.The Research of Chinese License Plates Recognition Based on CNN and Length Feature[M]∥Trends in Applied Knowledge-Based Systems and Data Science.Springer International Publishing,2016:389-397.
[9]AHRANJANY S S,RAZZAZI F.A Very High Accuracy Handwritten Character Recognition System forFarsi/Arabic Digits Using Convolutional Neural Networks[C]∥Liverpool:International Conference on Bio-Inspired Computing.2010:1585-1592.
[10]赵志宏,杨绍普,马增强.基于卷积神经网络LeNet-5的车牌字符识别研究[J].系统仿真学报,2010,22(3):638-641.
[11]张立,朱玉全,陈耿.基于卷积神经网络SLeNet-5的车牌识别方法[J].信息技术,2014,11:7-11.
[12]LIU Y J,HUANG H.Car Plate Character Recognition Using a Convolutional Neural Network with Shared HiddenLayers[C]∥Chinese Automation Congress.Wuhan,2015:638-643.
[13]YUAN A Q,BAI G,JIAO L J.Offline Handwritten English Character Recognitionbased on Convolutional Neural Network[C]∥International Workshop on Document Analysis Systems.Toronto,2012:125-129.
[14]KRIZHEVSKY A,SUTSKEVER I,HINTON G E.ImageNet classification with deep convolutional neural networks[C]∥International Conference on Neural Information Processing Systems.Curran Associates Inc.2012:1097-1105.
[15]LIN M,CHEN Q,YAN S C.Network In Network[D].Banff:International Conference on Learning Representations,2014:1-10.
[16]刘万军,梁雪剑,曲海成.不同池化模型的卷积神经网络学习性能研究[J].中国图象图形学报,2016,21(9):1178-1190.
[17]郭荣艳,胡雪惠.BP神经网络在车牌字符识别中的应用研究[J].计算机仿真,2010(9):299-301.