基于最近邻量化距离聚类的残差中心聚合图像表示

Residual Central Aggregation Image Representation Based on Nearest Neighbor Quantization Distance Clustering

通过累积残差和进行图像表示的局部聚合描述子向量(Vector of locally aggregated descriptors,VLAD)方法中,由于每个描述子与对应的最近邻码字得到的残差值大小不一,且每个码字对应的描述子数量不确定,会存在过累积和欠累积问题。针对此问题,提出一种通过距离聚类的残差中心聚合进行图像表示的新方法。首先,提取数据库图像的局部描述子,通过聚类得到码本;然后,将局部描述子通过最近邻方法量化到码本上,并求出局部描述子与最近邻码字之间的欧式距离;再次,聚类所有距离,得到中心集合,求出每个局部描述子与最近邻码字之间的欧式距离在中心集合上的最近邻,进而求得中心集合中每个中心对应的描述子与最近邻码字之间残差的中心,并将每个码字上所有的残差中心累积求和;最后,将所有码字对应的累积向量按顺序级联后得到最后的图像表示。在Holidays和UKB数据集上的图像检索实验结果表明,提出的图像表示方法比通过直接累积残差和进行图像表示的VLAD方法效果更好。

In the vector of locally aggregated descriptors(VLAD)method for image representation by residual accumulation,the residual values obtained by each descriptor and the corresponding nearest neighbor codeword are different,and the number of descriptors corresponding to each codeword is uncertain.Thus,there are over cumulative and under cumulative problems.In this paper,a new method for image representation by residual center aggregation of distance clustering is proposed.First,the local descriptors of the database image are extracted,and the codebook is obtained by clustering these descriptors;then,the local descriptors are quantized to the codebook by the nearest neighbor method,and the Euclidean distances between the local descriptors and the corresponding nearest neighbor codeword are obtained.Again,after clustering all the distances and obtaining the central set,the method finds the nearest neighbor of the Euclidean distance between each local descriptor and the nearest neighbor codeword on the central set,obtains the descriptor corresponding to each center in the central set,determines the center of the residual between the descriptor corresponding to each center in the center set and the nearest neighbor codeword,and accumulates and summarizes all the residual centers on each codeword.Finally,all the cumulative vectors corresponding to the codewords are cascaded in order to get the final image representation.The results of image retrieval experiments on the Holidays and UKB datasets show that the proposed image representation method is better than the VLAD method by directly accumulating residuals and performing image representation.