This letter presents a two-stage statistical approach for ``exploring and explaining'' a pictorial database, for content-based image retrieval systems. First, we describe how correspondence analysis provides images classes, as well as facilitates the understanding of the role of image primitives and attributes used to index pictures. Such understanding allows an intelligent choice of features, and thus computational savings, to be made. Second, ascendant hierarchical classification permits the structuring of the database, in order to ease picture indexing and retrieval.