基于DCNN特征与集成学习的车型分类算法  

Vehicle classification algorithm based on DCNN features and ensemble learning

作　　者：李大湘 王小雨[1,2] LI Da-xiang;WANG Xiao-yu(College of Communication and Information Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;Ministry of Public Security Key Laboratory of Electronic Information Application Technology for Scene Investigation,Xi’an University of Posts and Telecommunications,Xi’an 710121,China)

机构地区：[1]西安邮电大学通信与信息工程学院,陕西西安710121 [2]西安邮电大学电子信息现场勘验应用技术公安部重点实验室,陕西西安710121

出　　处：《计算机工程与设计》2020年第6期1624-1628,共5页Computer Engineering and Design

基　　金：陕西省国际合作交流基金项目(2017KW-013、2019JM-604);国家自然科学基金项目(61571361、61102095);西安邮电大学研究创新基金项目(CXJJLY2018037)。

摘　　要：针对传统人工设计特征描述不充分及单分类器泛化能力弱等问题,提出一种基于深度卷积神经网络(DCNN)特征与集成学习相结合的车型分类算法。微调VGG16深度卷积神经网络模型,将全连接层Fc7输出的4096维矢量采用PCA方法降至100维,作为图像的特征表示;采用拉格朗日支持向量机(LSVM)作为基分类器,以Adaboost方法自动学习各样本及基分类器的权重实现分类器集成。基于BIT和MIO-TCD数据集的对比实验结果表明,平均分类精度分别达到84.5%与83%,优于其它传统特征与单分类器方法。Aiming at the problems of inadequate feature description of traditional manual design and weak generalization ability of single classifier,a vehicle classification algorithm based on deep convolutional neural network(DCNN)feature and ensemble learning was proposed.VGG16 deep convolutional neural network model was fine-tuned,and PCA method was adopted to reduce the 4096-dimension vector outputted by Fc7 of the full connection layer to 100-dimension,and the results were taken as the feature representation of the image.Lagrange support vector machine(LSVM)was adopted as the base classifier,and Adaboost method was used to automatically learn the weights of each sample and base classifier to realize the classifier integration.Experimental results based on BIT and MIO-TCD data sets show that the average classification accuracies reach 84.5%and 83%respectively,which are better than that of other traditional features and single classifier methods.

关 键 词：深度卷积神经网络 集成学习 车型分类 拉格朗日支持向量机 提升算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]