融合FV-SIFT特征和深度卷积特征的车辆图像细粒度分类  

Vehicle fine-grained classification based on FV-SIFT feature and deep convolution feature

作　　者：杨志钢 马俊杰 YANG Zhigang;MA Junjie(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院

出　　处：《应用科技》2019年第4期42-47,共6页Applied Science and Technology

基　　金：国家自然科学基金项目(61201238)

摘　　要：针对现有的SIFT特征在车辆细粒度分类中存在的分类精度低的问题,提出了一种融合FV-SIFT特征和深度卷积特征的车辆图像细粒度分类算法。首先采用SIFT算法与Fisher Vector算法相结合的方式提取车辆图像的FV-SIFT特征,然后采用VGG-16卷积神经网络提取车辆图像的深度卷积特征,最后将FV-SIFT特征与深度卷积特征进行线性融合并采用支持向量机对融合后的车辆特征进行分类。实验结果表明,该方法的分类准确率达到82.3%,较FV-SIFT算法在分类准确率上提高了15.4%。For the low classification accuracy by SIFT feature in vehicle fine?grained classification,this paper pro?poses a novel vehicle fine?grained classification algorithm based on FV-SIFT and deep convolutional features.First?ly,the FV-SIFT features of vehicle image were extracted by combining SIFT algorithm and Fisher Vector algorithm.Then,the deep convolutional features of vehicle image were extracted by VGG-16 convolutional neural network.Fi?nally,the FV-SIFT features and the deep convolution features were linearly fused and the fused features were clas?sified by Support Vector Machine.Experimental results show that the classification accuracy of this method is 82.3%,which is 15.4%higher than the FV-SIFT algorithm.

关 键 词：图像细粒度分类 SIFT算法 Fisher Vector算法 卷积神经网络 SVM分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]