Visual information retrieval from 2D shapes by bipolar-matching

One of the most challenging issues in visual information retrieval is retrieval by shape, due to a lack of mathematically rigorous definition of shape similarity. This paper presents a bipolar model for computing shape similarity. Given a discrete region, we cut its Voronoi diagram into two parts along the border of the region and each part is a tree. We use the two trees to respectively model the structures of a region and its complement, which is called the Bipolar Model. We prune the two trees by removing the nodes with small protrusions. The leaf nodes of the pruned trees are interleaved to make a leaf chain. Two regions are compared and matched, using a cyclic edit distance between the two leaf chains, with restricted merge and split operations allowed. We tested our algorithm on the MPEG-7 data set and made a “bullseye” score of 89.9%, which is the best performance ever reported.