A filter and refinement approach to RBIR

A promising trend in content based image retrieval (CBIR) is the incorporation of the notion of objects into the similarity evaluation. Images are automatically segmented into a dynamic number of regions that roughly correspond to objects. In region based image retrieval (RBIR), we compute standard feature characteristic atop those segments and evaluate the similarity between images based on the similarity of its segments. This requires the application of matching techniques, like the Hungarian algorithm, that are prohibitivly expensive for collections beyond a few thousand images. In this paper, we provide a filter and refinement search algorithm to solve RBIR queries in reasonable times for even large image collections. We describe the bounding functions as well as implementational issues of our approach. Experimental results show that retrieval performance surpasses alternative approaches by a factor of 5.