基于改进VLAD算法的图像分类  

Image classification based on improved VLAD

作　　者：王倩[1] 张新明 蔡强[1] 祝晓斌[1] Wang Qian;Zhang Xinming;Cai Qiang;Zhu Xiaobin(School of Computer&Information Engineering,Beijing Technology&Business University,Beijing 100048,China)

机构地区：[1]北京工商大学计算机与信息工程学院,北京100048

出　　处：《计算机应用研究》2018年第10期3151-3154,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(61402023)

摘　　要：针对图像分类问题进行了研究,提出一种改进的局部聚合描述符(vector of locally aggregated descriptors,VLAD)算法以得到高效的图像特征表示。采用卷积神经网络提取图像的密集局部特征。正态分布式选取子集训练视觉字典,提高字典质量;然后,采用多近邻分配代替最近邻匹配,将特征量化到多个视觉字典且赋予不同的权重;最后,基于VLAD原理对图像局部特征进行编码,并用支持向量机对目标进行分类。在多个数据集上的实验结果表明,与近年提出的几种经典的图像分类算法相比,所提方法取得了较高的分类正确率。In view of the research of image classification,this paper proposed an improved VLAD algorithm for efficient image representation.Firstly,it extracted the dense local features of images based on the trained convolutional neural networks.Se-condly,it selected a subset of features by the criterion of normal distribution for high quality codebook generation.Thirdly,it assigned the local features to multi-neighbor visual words in the codebook instead of the nearest one with different weights.Finally,it encoded the local features based on the principles of VLAD,and adopted SVM for object classification.Experimental results show that the proposed method can achieve better performance than the state-of-the-art image classification methods.

关 键 词：局部聚合描述符 图像分类 卷积神经网络 特征编码 多近邻分配 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]