In this a novel supervised learning method is proposed to map low-level visualfeatures to high-level semantic conceptsfor region-based image retrieval. The contributions of thispaper lie in threefolds. ( I ) For each semantic concept, a set of low-level tokens are extracted fmm the segmented regions of training images. Those tokens capture the representative informationfor describing the semantic meaning of that concept; (2)A set of posteriors are generated based on the low-level tokens through which denote the probabilities of images belonging to the semantic concepts. Theposteriors are treated as high-levelfeatures that connect images with high-level semantic concepts. Long-term relevance feedback learning is incorporated to provide the supervisory information needed in the above learning process, including the concept informationand the relevanttraining setfor each concept; (3)An integratedalgorithm is implemented to combine two kinds of informationfor retrieval: the informationfro...