Finding an image object-modeling paradigm that is robust, flexible, and highly adaptive is essential for content-based image retrieval and object recognition. We model the image objects by attributed feature-relational graphs (AFRG), in which the nodes are the features, including all kinds of feature points and regions, the arcs are the edges connecting the feature points and the vicinity relationship between regions. Both nodes and edges have their specific set of attributes. We developed algorithms to extract these features in the graph from color images. Then we group these features into perceptual organizations that have certain special relationship and connectivity. Because the type and attributes of the nodes and perceptual organizations them have strong saliencies and are invariant to different levels of viewpoint and lighting changes, these perceptual organizations capture a great amount of the essential information of the image objects. We developed a hypothesis-verification strategy to match the AFRGs based on the perceptual organizations contained in them. We also developed an experimental image retrieval system to show that our approach is very effective and have great potentials. |