Ontology driven content based image retrieval

Content based image retrieval (CBIR) methods are proposed as alternative or complementary solutions to keyword-based picture search. However these techniques mostly rely on low-level descriptors similarity between different items and when one uses such an application to find pictures, the proposed answers are often not conceptually similar to the query. In this paper, we describe RetrievOnto, an image retrieval (IR) system that allies CBIR techniques and semantics in order to better fit the users' expectations when querying an image database. The dataset is structured employing a term hierarchy, which is used to control the conceptual neighbourhood where similar items are searched. Only the leaf terms of the hierarchy have associated image sets but, with the use of the type-subtype relation between nodes, pictures are indirectly associated to all the concepts in the hierarchy and the system can propose localized IR processes, which associate low-level and conceptual similarities (on different levels of generality). We model a real-world situation by using pictures gathered from the Internet. The ontologically controlled IR method proposed in this paper is compared to classical CBIR functioning and we show that the introduction of a hierarchical structure improves precision results for the system.