A challenging problem in image retrieval is the combination of multiple features and similarity models. We pose the retrieval problem in a two-level classification framework with two classes: the relevance class and the irrelevance class of the query. The fir st level maps high-dimensional feature spaces to two-dimensional probability spaces. The second level uses combinations of simple linear classifiers trained in these multiple probability spaces to compensate for errors in modeling probabilities in feature spaces. Similarity is computed using joint posterior probability ratios instead of the common way of computing distances in feature spaces and taking their weighted combinations. Experiments on two groundtruthed databases show that the proposed classification framework performs significantly better than the common geometric framework of distances and allows a well-defined and effective way of combining multiple features and similarity measures.