Content-Based Image Retrieval through Aggregation of Similarity Measures

Abstract

A multiple (dis)similarity aggregation framework is proposed for robust selection of suitable measures for a query. Ranges of all measures are normalized and transformed into a single similarity measure by re-scoring them. Precision levels and computational costs on the Corel database are compared to a system that uses a single measure. The combination of Minkowski distance and cosine similarity achieves the best result in the sense that its computational overhead is only 0.1 seconds when 100,000 images and 1,000 visual words are used. Due to the robustness of measures in this combination, applications include retrieval employing large number of image categories in a database.