This paper describes an approach to personalized image indexing and retrieval. To tackle the issue of subjectivity in Content-Based Image Retrieval (CBIR), users can define their own indexing vocabulary and make the system learn it. These indexing concepts may be both local (objects) and global (image ategories). The system guides the user in the selection of relevant training examples. Concept learning in the system is incremental and hierarchical: global concepts are built upon local concepts as well as low-level features. Similarity measures tuning is used to emphasize relevant features for a given concept. To illustrate the potential of this approach, an implementation of this model has been developed; preliminary results are given in this paper.