In this paper, the authors propose a kernel-based approach to improve the retrieval performances of CBIR systems by learning a distance metric based on class probability distributions. Unlike other metric learning methods which are based on local or global constraints, the proposed method learns for each class a nonlinear kernel which transforms the original feature space to a more effective one. The distances between query and database images are then measured in the new space. Experimental results show that the kernelbased approach not only improves the retrieval performances of kernel distance without learning, but also outperforms other kernel metric learning methods.