Image Indexing and Retrieval Using Visual Terms and Text-Like Weighting

Abstract

Image similarity is typically evaluated by using low level features such as color histograms, textures, and shapes. Image similarity search algorithms require computing similarity between low level features of the query image and those of the images in the database. Even if state of the art access methods for similarity search reduce the set of features to be accessed and compared to the query, similarity search has still an high cost.

In this paper we present a novel approach which processes image similarity search queries by using a technique that takes inspiration from text retrieval. We propose an approach that automatically indexes images by using visual terms chosen from a visual lexicon.

Each visual term represents a typology of visual regions, according to various criteria. The visual lexicon is obtained by analyzing a training set of images, to infer which are the relevant typology of visual regions. We have defined a weighting and matching schema that are able respectively to associate visual terms with images and to compare images by means of the associated terms.

We show that the proposed approach do not lose performance, in terms of effectiveness, with respect to other methods existing in literature, and at the same time offers higher performance, in terms of efficiency, given the possibility of using inverted files to support similarity searching.

This work was partially supported by the DELOS NoE and the Multimatch project, funded by the European Commission under FP6 (Sixth Framework Programme).