Cross-Resolution Image Similarity Modeling

A cross-resolution image similarity model employing probabilistic interpretations of the similarity values turned into fuzzy values combined with suitable merge functions is presented. Masking the negative particularities of individual region-based image similarity models, five region-based image similarity models are used at the same time when calculating the overall image similarity. By employing aggregation operators, capturing of a variety of conjunctive, disjunctive, and other non-linear combinations of similarity criteria is allowed. Empirical evaluation of the proposed model on four test databases, containing 4,444 images in 150 semantic categories was carried out. The results obtained from the evaluation revealed that cross-resolution image similarity modeling results in optimal retrieval performance. Compared to two well-known image retrieval systems, SIMPLicity and WBIIS, the proposed model brings an increase of 1.7% and 22% respectively in average retrieval precision. The experimental evaluation presented may thus be helpful and suggest possible further improvements can be achieved along the same line of research directions in various computer vision tasks.