基于内积加权局部聚合描述子向量的图像分类

Image classification based on inner product weighted vector of locally aggregated descriptors

局部聚合描述子向量(vector of locally aggregated descriptors,VLAD)是一种硬编码方式,会导致较大的量化损失。为了解决此问题,提出了一种基于内积加权的VLAD编码(inner product weighted vector of locally aggregated descriptors,IPWVLAD),它是一种软编码方式,为图像中的每个描述子寻找若干个近邻的基向量,并采用内积编码的方式生成权重信息添加到累积残差中。对于最近邻的基向量和描述子之间的残差给予最大的权重,对于次近邻的情况依次赋予越来越小的权重。在Corel 10、15 Scenes、UIUC Sport Events数据集上的实验结果表明,与已有的4种基于VLAD的方法和2种常用的表示方法相比,本文所提出的IPWVLAD编码获得了较好的分类性能。

The vector of locally aggregated descriptors(VLAD)is a hard encoding method,which would lead to large quantization loss.In order to solve this problem,an inner product weighted vector of locally aggregated descriptors(IPWVLAD)was proposed,which was a soft encoding method.It searched for several neighboring base vectors for each descriptor in the image,and used inner product encoding method to generate weight information,which was added to the accumulated residual.The residual between the nearest neighbor’s base vector and the descriptor was given the maximum weight,and the next neighbor’s case was given the smaller and smaller weight in turn.The experimental results on Corel 10,15 Scenes and UIUC Sport Events datasets show that the proposed IPWVLAD encoding has good classification performance compared with the existing four methods based VLAD and two commonly used representation methods.