Buoy Indexing of Metric Feature Spaces for Fast Approximate Image Queries

Abstract

A novel indexing scheme for solving the problem of nearest neighbor queries in generic metric feature spaces for content-based image retrieval is proposed to break the 'dimensionality curse.' The basis for the proposed method is the partitioning of the feature dataset into clusters that are represented by single buoys. Upon submission of a query request, only a small number of clusters whose buoys are close to the query object are considered for the approximate query result, effectively cutting down the amount of data to be processed enormously. Results concerning the retrieval accuracy from extensive experimentation with a real image archive are given. The influence of control parameters is investigated with respect to the tradeoff between retrieval accuracy and computational cost.