基于交叉熵的CNN交通标志识别方法研究
2018-02-12陈玉婷张凤瑜
陈玉婷 张凤瑜



摘要:为了弥补经典LeNet-5卷积神经网络模型在交通标志识别中易发生过拟合的不足,克服其识别准确率与训练效率较低的缺点,采用GTSRB德国交通标志数据集,并对数据集进行扩充,使训练集数量达到6 312 649张,同时对经典LeNet-5卷积神经网络模型在激活函数、池化策略、随机丢弃及网络结构等方面进行改进,使用基于交叉熵的梯度下降算法对模型的误差反向传播过程进行优化。实验结果表明,改进后的模型较改进前在识别准确率与训练效率上都有所提高,识别准确率最高可达97.04%。因此,基于交叉熵的卷积神经网络不仅能够提高模型的交通标志识别准确率,有效防止过拟合,还能够提升网络训练效率。
关键词:交通标志识别;卷积神经网络;交叉熵
Research on CNN Traffic Sign Recognition Method Based on Cross Entropy
CHEN Yu ting, ZHANG Feng yu
(School of Information Science & Technology, Northeast Normal University, Changchun 130117,China)
Abstract:In order to make up for the shortcomings of the classic LeNet 5 convolution neural network model in traffic sign recognition and overcome the shortcomings of the recognition accuracy and low training efficiency, the data set of GTSRB German traffic sign is used and the data set is expanded to make the number of training setup to 6 312 649. The classical LeNet 5 convolution neural network model is improved in the aspects of activation function, pooling strategy, random discarding and network structure, and the algorithm of gradient descent based on cross entropy is used to optimize the error back propagation process of the model. Through the related experiments, the improved model has improved the recognition accuracy and training efficiency compared with the improvement, and the recognition accuracy reaches 97.04%. The convolution neural network based on cross entropy can not only improve the recognition accuracy of traffic signs recognition, effectively prevent overfitting, but also improve the training efficiency of the network.
Key Words:traffic sign recognition;convolution neural network;cross entropy
0 引言
近年来,智能交通决策系统得到了迅速发展,其能够识别交通标志信息并作出下一步决策,从而辅助甚至代替驾驶员驾驶,减轻驾驶员的驾驶压力,同时可对驾驶员的违规驾驶行为作出提醒,进而减少交通事故发生[1]。交通标志在道路交通中扮演着重要角色,其能够直接指导驾驶员在行车过程中的行为。目前,汽车上普遍安装了行车记录仪,可以通过其采集道路交通信息,获取交通标志,并进一步识别出交通标志的所属种类。随着城市的发展,道路交通安全设施日趋完善,交通标志的种类也日益丰富[2],因此对交通标志识别的泛化能力也有着较高要求。
随着机器学习、深度学习[3]等人工智能技术的不断发展与应用,在很多方面都取得了较好的应用效果,也有越来越多学者将卷积神经网络算法引入到识别问题中。……
