基于优化的卷积神经网络在交通标志识别中的应用  

Application of optimized convolutional neural network in traffic sign recognition

作　　者：张邯 罗晓曙[1] 袁荣尚 ZHANG Han;LUO Xiaoshu;YUAN Rongshang(College of Electronic Engineering,Guangxi Normal University,Guilin 541004,China)

机构地区：[1]广西师范大学电子工程学院,广西桂林541004

出　　处：《现代电子技术》2018年第21期132-136,共5页Modern Electronics Technique

基　　金：国家自然科学基金(11262004);广西多源信息挖掘与安全重点实验室开放基金(MIMS15-06);广西信息科学实验中心基金(KA1430);广西研究生教育创新计划项目(XYCSZ2017051)~~

摘　　要：在现实的交通环境中,由于各种因素影响,使得所采集到的交通标志图像识别的准确性不高,鲁棒性也较差,给交通标志的准确识别带来了很大的困难。为此,采用非对称卷积结构对经典卷积神经网络AlexNet进行改进,并引入批量归一化(BN)方法,提出基于优化卷积神经网络结构的交通标志识别方法。其中,非对称卷积结构使网络进一步加深,提高了识别精度。BN将每一层的输出数据归一化为均值为0、标准差为1,确保了数据稳定,使梯度传输更为顺畅。使用德国交通标志数据集进行训练并测试,结果显示改进的网络结构提升了网络的分类精度,且达到了97.56%,具有一定的应用价值。In the real traffic environment,the recognition of collected traffic sign images has low accuracy and poor robustness affected by various factors,which brings a great difficulty in the accurate identification of traffic signs.Therefore,the asymmetric convolution structure is used to improve the classical convolutional neural network(CNN)Alexnet,and the batch normalization(BN)method is introduced to propose the traffic sign recognition method based on optimized CNN structure.The asymmetric convolution structure further optimizes the CNN network and improves its recognition accuracy.The output data of each layer is normalized into a mean value 0 and a standard deviation 1 by means of BN,which can ensure the stability of data and makes the gradient transmission more smooth.The data set of German traffic sign is used for training and recognition.The results show that the improved network structure improves the classification accuracy of the network,and its classification accuracy can reach up to 97.56%,which has a certain application value.

关 键 词：卷积神经网络 非对称卷积 批量归一化 交通标志 梯度传输 分类精度 

分 类 号：TN711-34[电子电信—电路与系统] TP391.4[自动化与计算机技术—计算机应用技术]