摘要
研究车身颜色识别技术可为公安部门打击套牌车辆等交通犯罪行为提供技术支持,并为道路行车安全提供理论依据。文章基于Caffe深度学习框架,提出了一种基于深度卷积神经网络车身颜色识别技术的研究方案,分析了网络层数、迭代次数和学习率对模型的影响,对卷积神经网络CNN模型进行优化,并将优化的网络模型与支持向量机SVM、改进后的HSV模型进行对比分析。结果表明:卷积神经网络最优模型的神经网络层数为8,最大迭代次数为30万,学习率为0.001;支持向量机SVM、改进后的HSV颜色模型及卷积神经网络识别率分别为80.05%、85.25%、90.20%。
Research on vehicle body color recognition technology can provide technical support for public security departments to crack down on traffic offenses such as decks and vehicles,and can provide a theoretical basis for road traffic safety. Based on caffe deep learning framework,this paper proposes a research program of vehicle body color recognition based on deep convolution neural network. The influence of network layer number,iteration number and learning rate on the model is studied. CNN model of convolution neural network is optimized,and the optimized network model is compared with support vector machine SVM and improved HSV model. The results show that the neural network optimal layer model of convolutional neural network has a number of layers of 8,a maximum number of iterations of 300,000 and a learning rate of 0. 001. SVM recognition rate of support vector machine is 80. 05%,and recognition rate of improved HSV color model is 85. 25 %,Convolutional neural network recognition rate of 90. 20%,proving that convolution neural network in the recognition of the best.
作者
管德永
鞠铭
安连华
Guan Deyong;Ju Ming;An Lianhua(School of Traffic, Shandong University of Science and Technology, Qingdao 266590, Chin)
出处
《山东建筑大学学报》
2018年第1期25-31,共7页
Journal of Shandong Jianzhu University
基金
山东科技大学人才引进科研启动基金资助项目(2015RCJJ032)
关键词
卷积神经网络
车身颜色识别
网络层数
迭代次数
学习率
convolutional neural network
vehicle color recognition
network layer number
iteration number
learning rate
