基于改进LeNet-5网络的车牌字符识别  

License Plate Character Recognition Based on Improved LeNet-5 Model

作　　者：张秀玲[1] 魏其珺 周凯旋 董逍鹏 马锴[1] ZHANG Xiuling;WEI Qijun;ZHOU Kaixuan;DONG Xiaopeng;MA Kai(Key Laboratory of Industrial Computer Control Engineering of Hebei Province,Yanshan University, Qinhuangdao, 066004, China)

机构地区：[1]燕山大学河北省工业计算机控制工程重点实验室,河北秦皇岛066004

出　　处：《沈阳大学学报（自然科学版）》2020年第4期312-317,共6页Journal of Shenyang University：Natural Science

基　　金：国家自然科学基金资助项目(61573303)。

摘　　要：引入了Inception-SE卷积模块组来提升LeNet-5网络的广度与深度,运用SE模块增强了有用的特征并抑制了对当前任务用处不大的特征;使用BN层和Dropout优化网络,防止梯度弥散,提升精度;使用全局池化层(global average pooling,GAP)代替全连接层来减少网络计算参数.研究结果表明:改进后网络的识别精度达到了99.88%,比传统的LeNet-5网络提高了1.71%.The Inception-SE convolution module group was introduced to enhance the breadth and depth of the LeNet-5 network,and the use of the SE module enhanced useful features and suppressed features that are not very useful for the current task.The BN layer and Dropout were used to optimize the network,prevent the gradient dispersion and improve the recognition rate.The global average pooling(GAP)was used to replace the full connection layer to reduce the network calculation parameters.The research results show that the recognition rate of the improved network is 99.88%,and the recognition accuracy is 1.71%higher than that of the traditional LeNet-5 network.

关 键 词：卷积神经网络 车牌字符识别 LeNet-5网络 Inception-SE卷积模块 识别精度 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]