ASIA: An Automatic Annotation Technique for Query-by-Concept in Image Retrieval Systems

In this paper we introduce ASIA, a sampling-based image annotation technique for high-level image retrieval. ASIA consists of two phases. In the first phase we create off-line an annotated image database with high-level image concepts using low-level image features such as the colors and structure of sample objects. This annotation technique uses a new sampling-based matching framework relatively more robust to rotation, scaling and translation invariants. In the second phase, the user can query the annotated image database using text-only concepts. We show, by experimental results, that the ASIA method is significantly better than a recently proposed alternative Monotonic Tree method. ASIA has the ability to identify more relevant images, while producing fewer false matches.