Cross-domain image matching improved by visual attention

Abstract

A good accuracy in image retrieval across different visual domains, such as photos taken over different seasons or lighting conditions, paintings, drawings, hand-drawn sketches, still is a big challenge. This paper proposes the use of visual attention to estimate the relative importance of some regions in a given query image. Recently, researchers used different databases in specific domains to validate their hypothesis. In this paper, we also propose a database with multiple image domains, called UFU-DDD. We used the UFU-DDD database to demonstrate the performance and accuracy gains from the association of visual attention with orientation-based feature descriptors. The analysis of the results showed that our approach outperforms all the standard descriptors used in the experiments. We hope the UFU-DDD database constitutes a valuable benchmark to the future research in cross-domain similarity searching.