A system for effective content based image retrieval

Image retrieval based on semantic contents aims to facilitate intuitive and effective retrieval for the unsophisticated user. We demonstrate a system which employs several techniques to maximize retrieval effectiveness while providing user friendly querying. Our system consists of an icon-based visual query tool with a dynamic object/attribute library and a search engine. The users construct their queries graphically by dragging icons (representing physics! objects) into a window called the “Canvas”, describing the attributes by clicking on these icons and optionally describing their relationships. We employ the following techniques for effective image retrieval: Normalization, object significance, conflict elimination, and restricted matching across types.