Journal of Frontiers of Computer Science and Technology ›› 2018, Vol. 12 ›› Issue (2): 282-291.DOI: 10.3778/j.issn.1673-9418.1704055

Previous Articles     Next Articles

Application of Convolutional Neural Network in Vehicle Recognition

PENG Qing1, JI Guishu1+, XIE Linjiang1, ZHANG Shaobo1,2   

  1. 1. School of Information Science and Engineering, Central South University, Changsha 410083, China
    2. School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan, Hunan 411201, China
  • Online:2018-02-01 Published:2018-01-31

卷积神经网络在车辆识别中的应用

彭  清1,季桂树1+,谢林江1,张少波1,2   

  1. 1. 中南大学 信息科学与工程学院,长沙 410083
    2. 湖南科技大学 计算机科学与工程学院,湖南 湘潭 411201

Abstract: Aiming at the problems of excessive calculation and complex feature extraction of existing vehicle recognition methods, this paper proposes a vehicle recognition method based on convolutional neural network (CNN). Firstly, this paper constructs a convolutional neural network model, which is trained with different size of convolution kernel, different number of network layers and different number of feature maps. Secondly, this paper obtains the optimal model through 100 iterations learns, from which to extract all features of hidden layer and combined with support vector machines (SVM) to proceed with recognition. Finally, this paper systematically analyzes the influence of different parameters on the accuracy and mean square error. The experimental results show that in vehicle recognition CNN+SVM had a high accuracy rate as compared to the traditional CNN, PCA+SVM, HOG+SVM and Wavelet+SVM, whose accuracy rate is 97.00%. This paper focuses on analyzing the cause for errors in samples and necessary modifications to be done hereafter.

Key words: vehicle recognition, deep learning, convolutional neural network (CNN), feature extracting, support vector machine (SVM)

摘要: 针对现有车辆识别方法计算量大,提取特征复杂等问题,提出一种基于卷积神经网络(convolutional neural network,CNN)的车辆识别方法。构建卷积神经网络模型,分别使用不同的卷积核、网络层数、特征图数对网络进行训练;通过100次迭代的学习结果得到最优模型,提取隐含层所有特征,并结合支持向量机进行识别;系统分析了不同参数对测试正确率和样本均方误差的影响。实验结果显示,CNN+SVM在车辆识别中的准确率明显优于传统CNN、PCA+SVM、HOG+SVM、Wavelet+SVM,正确率为97.00%,分析了样本识别错误的原因以及今后需要改进的地方,为以后的研究指明了方向。

关键词: 车辆识别, 深度学习, 卷积神经网络(CNN), 特征提取, 支持向量机(SVM)