A structured vocabulary of terms, such as a textual thesaurus, provides a way to conceptually describe visual information. The retrieval model described in this paper combines a conceptual and a visual layer as a first step towards the integration of ontologies and content-based image retrieval. Terms are related with image regions through a weighted association. This model allows the execution of concept-level queries, fulfilling user expectations and reducing the so-called semantic gap. Region-based relevance feedback is used to improve the quality of results in each query session and to help in the discovery of associations between text and image. The learning mechanism, whose function is to discover existing term-region associations, is based on a clustering algorithm applied over the features space and on propagation functions, which acts in each cluster where new information is available from user interaction. This approach is validated with the presentation of promising results obtained using the VOIR - Visual Object Information Retrieval system.