ISSN: 1673-825X    Imprint: Chongqing University of Posts and Telecommunications Journal
Volume listings
  • Current volume
  •  
  • Search by keywords:
  • Search by 
  • Keywords
  • To
  • Journal archive
Related Links
基于集成卷积神经网络的交通标志识别
Traffic sign recognition based on ensemble convolutional neural network
DOI:10.3979/j.issn.1673-825X.2019.04.019
Received:January 24, 2018  Revised:March 01, 2019
中文关键词:卷积神经网络  集成学习  批量归一化  支持向量机
英文关键词:convolutional neural network  ensemble learning  batch normalization  support vector machine
基金项目:国家自然科学基金(61571071)
Author NameAffiliationE-mail
ZHANG Gongguo School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, P. R. China zhanggongguo@cqcyit.com 
WU Jian School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, P. R. China wujian@cqupt.edu.cn 
YI Yi School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, P. R. China 752193664@qq.com 
WANG Ziquan School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, P. R. China  
SUN Haixia School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, P. R. China  
Hits: 45
Download times: 23
中文摘要:
      以交通标志识别为研究目的,提出一种基于集成卷积神经网络的交通标志识别算法,通过对多个不同结构的卷积神经网络进行集成以提高算法识别率。为了缩短网络训练和测试时间以及提高网络识别率,对单个卷积神经网络的结构进行了优化。使用ReLU(rectified linear unit)激活函数来代替传统的激活函数,使用批量归一化(batch normalization,BN) 方法对卷积层输出数据进行归一化处理,将卷积神经网络的分类器用支持向量机(support vector machine,SVM)代替。使用德国交通标志识别数据库(german traffic sign recognition benchmark,GTSRB)进行训练和测试,实验结果表明,提出的算法识别率为98.29%,单幅交通标志图像测试时间为1.32 ms,对交通标志具有良好的识别性能。
英文摘要:
      This paper presents a traffic sign recognition algorithm based on ensemble convolution neural network to study traffic identification. The multiple convolutional neural networks of different structures are integrated to improve the algorithm recognition rate. The structure of a single convolutional neural network is optimized to shorten the network training, test time and improve the network recognition rate. First, the rectified linear unit (ReLU) activation function is used to replace the traditional activation function. Then, batch normalization (BN) method is used to normalize the output data of convolution layer. Finally, the classifier of convolution neural network is replaced by SVM. This paper uses the German traffic sign recognition benchmark (GTSRB) for training and testing, and the experimental results show that the algorithm recognition rate is 98.29% and the image test time of single traffic sign is 1.32 ms, which has good recognition performance for traffic signs.
HTML    View PDF   View/Add Comment  Download reader