Relevance feedback (RF) and region-based image retrieval (RBIR) are two widely used methods to enhance the performance of contentbased image retrieval (CBIR) systems. We combine these two methods in this paper, and propose a region weighting scheme to refelct the process of human visual perception. Furthermore, rather than using a single positive feedback group, the proposed approach introduces RBIR to the relevance feedback with multiple positive and negative groups. To guide users in grouping the positive feedbacks, the proposed system provides a heuristic pre-clustering result automatically. Using these guiding clusters, the users can re-group the positive feedbacks to express his/her particular interests. Finally, Group Biased Discriminant Analysis (GBDA) is modified and applied to the similarity measure between images constructed on the basis of the regionbased relevance feedbacks.