ABI: analogy-based indexing for content image retrieval

Abstract Morphogeometric-based metrics are not always appropriate to describe high-level contents of images as well as to formulate complex queries. People often find that two pictures as similar because they share relational predicates rather than objects attributes. In particular, images can be related because they are analogous . Scientists, for example, use analogies to trace art influences across different paints. In this paper, we focus on analogous relationships between groups of objects. The model we propose combines primitive properties by mean of a logical reasoning engine to produce a hierarchical image description. Each picture is decomposed into its spatial relations (physical layer), cognitive relations between objects within a group (group layer), and relations between groups (meta-group layer). This new Analogy Based Indexing (ABI for short) for Content Image Retrieval, allows users to express complex queries such as search for functional associations or group membership relations. A proof-of-concept prototype is also discussed to verify the precision and the efficiency of the proposed system. Furthermore, an embedded visual language enables pictorial queries composition and simplifies image annotation. The experimental results show the effectiveness of ABI in terms of precision vs. recall curve diagrams.