Visual information retrieval based on shape similarity

An effective and fast shape description and retrieval method is presented for huge image databases. As a shape representation for deformable objects, a multi-scale skeleton representation is proposed in order to preserve the consistency of the skeletons and to reduce the effect of the structural changes. Incorrect matches due to the boundary noise in a segmentation process are avoided by including multiple coarse skeletons of different scales. A fast computational method for the similarity of skeletons is also proposed by using the moment invariants. Experimental results on animal databases showed that the proposed method gives prominent accuracy in retrieval.