Image Retrieval Using Sieve Complement Trees

This paper is about scale-space image trees. We introduce here a variant of the sieve algorithm to produce sieve complement trees where not only extremal regions are characterized but also their corresponding complements. Different simplification methods can transform (or prune) the hierarchy into a simpler form where the remaining nodes represent regions that are noticeably different from their neighbourhood, and that are traditionally known in the literature as "Salient Regions". Although, the resulting scale-space tree hierarchy, can not strictly be defined as a segmentation, its associated signal (a simplified image of the original), can be used for content based image retrieval (CBIR) similarly to a segmentation. Here, we present the resulting retrieval precision rates of testing our trees into three widely known image datasets. Our results confirm the premise that complementary regions can contribute to improve image retrieval rates.