Saliency distinguishing and applications to semantics extraction and retrieval of natural image

A system for image semantic information extraction and retrieval based on combination between top-down, target-driven and bottom-up, image-driven visual saliency distinguishing is presented. It simulates the analogous procedures of visual system of primates. It consists of four major components: selective attention, semantics extraction, semantics modeling and image retrieval. The design of selective attention unit includes two steps: selection of optimal features that optimally distinguish between objects and distracting background derived from top-down visual attention and statistic property of image, detecting desired salient objects based on integration between top-down and bottom-up visual attention mechanism. Semantics extraction unit performs the extraction of the object and background semantics of image in terms of classification decision rules. The semantics modeling unit is to construct the image semantic model. It produces a well-structured semantic representation to facilitate the visual information retrieval. Image retrieval component can perform the retrieval processes according to key words, desired objects, and instance of image respectively in order to gain efficient and adaptable retrieval performance. The experimental results show that the system introduced in this paper can provide effective adaptable and scalable image high-level semantic representations and efficient image retrieval performance.