Naming of image regions for user-friendly image retrieval

This paper presents a novel method for facilitating user-friendly image retrieval by attaching names to image regions. We first detect only the most prominent regions in images when such entities exist, using our own nonlinear image segmentation technique. Besides their visual features, the layout and relations between selected regions are also emphasized. Next, we apply an adaptive and multi-modal classification and naming of image regions using subsequent clustering methods to the features of the regions and related words as well as relevancy information. For both the naming and the testing, we have added a set of illustrations acting as abstract prototypes of the regions to randomly selected natural images. Experiments on 20,000 natural images show the efficacy of using this multilayer region naming model as well as of extensively interacting with users, enabling them to present their queries by a combination of region names, sketches and example images or regions.