Compact hierarchical image descriptors

Abstract

An efficient approach for image annotation and retrieval is presented. The main objective is to overcome the speed limitations of existing video indexing and retrieval systems. A family of successively simplified image descriptors is defined using colour histograms. The colour distribution is quantised according to a scale-parameter obtained by convolving the distribution function with Gaussians of increasing size. The generated family of histograms is then used to define descriptors at various levels of detail. Using this hierarchical descriptor structure large sets of non-similar images are discriminated at very low computational cost using low detailed descriptions. The search is then refined progressively until only a few very similar objects or images are found and ranked using higher levels of detail. To compare image descriptors the earth mover's distance is used. Experiments have been conducted to assess the performance of the proposed technique.