Geometric-Similarity Retrieval in Large Image Bases

We propose a novel approach to shape-based image retrieval that builds upon a similarity criterion which is based on the average point set distance. Compared to traditional techniques, such as dimensionality reduction, our method exhibits better behavior in that it maintains the average topology of shapes independently of the number of points used to represent them and is more resilient to noise. An efficient algorithm is presented based on an incremental ``fattening'' of the query shape until the best match is discovered. The algorithm uses simplex range search techniques and fractional cascading to provide an average poly-logarithmic time complexity on the total number of shape vertices. The algorithm is extended to perform additional fast approximate matching, when there is no image sufficiently similar to the query image. We present techniques for the efficient external storage of the shape base and of the auxiliary geometric data structures used by the algorithm. Finally, we show how our approach can be used for processing queries, containing pairwise relations of object boundaries such as contain, tangent, and overlap. Such queries are either extracted from some user drafted sketch or defined explicitly by the user. Alternative methods are presented for forming query execution plans.