A CBIR framework: Dimension reduction by radial basis function

In classical content-based image retrieval (CBIR) system, using Euclidean metric, it usually can not achieve good results, because of the semantic gap. To solve the difficulty problem, present a relevance feedback(RF) paradigm which is naturally guided only on dimension reduction with radial basis function(RBF). While images are often represented by feature vectors, the distance is usually different from the distance induced by the space. The geodesic distances on manifold are employed to measure the similarities between images. According to man interactions(loops) in a RF driven query-by-example system, the inhere similarities between images can be exactly estimated. We design a algorithm framework, in order to approximate the optimal mapping function with a RBF neural network(NN). The semantic gap (SG) of a new visual image can be inferred by the RBF neural network. Experiment results show that our method is effective in improving the performance of visual retrieval.