Fuzzy Multi-Resolution Image Similarity Modeling applied to Image Retrieval

Abstract

A multi-resolution image similarity model based on region-based image similarity modeling and fuzzy aggregation operators is presented, where the overall image similarity between two images is based on fuzzyin three sets of crisp valued similarities: feature, region and image, respectively, in a hierarchical manner. It helps reducing the influence of inaccurate image segmentations of the global and region-based image similarity models. Compared with the image similarity modeling on either global or region-based representation with crisp valued feature, region or image similarity representations, the proposed modeling results in the better overall retrieval performance with an average retrieval precision higher between 2% and 6%. Compared to two image retrieval systems, SIMPLicity and WBIIS, the proposed model brings an increase of 2% and 22% respectively in average retrieval precision. The descriptive power of the image similarity model increases by allowing model itself to capture the variety of the similarity criteria when compared to the conventional image similarity models. The proposed multiresolution image similarity modeling is thus more suited when approximating human perception of the image similarity in image retrieval.