Content based image retrieval based on a nonlinear similarity model

In this paper, we propose a new nonlinear paradigm to clustering, indexing and searching for content-based image retrieval (CBIR). The scheme is designed for approximate searches and all the work is performed in a transformed feature space. We first (1) map the input space into a feature space via a nonlinear map, (2) compute the top eigenvectors in that feature space, and (3) capture cluster structure based on the eigenvectors. We (4) describe each cluster with a minimal hypersphere containing all objects in the cluster, (5) derive the similarity measure for each cluster individually and (6) construct a bitmap index for each cluster. Finally we (7) model the similarity query as a hyper-rectangular range query and search the clusters near the query point. Our preliminary experimental results for our new framework demonstrate considerable effectiveness and efficiency in CBIR.