This work takes place within the framework of color image retrieval in specialized databases. A CBIR scheme has been proposed allowing to use a description thanks to low-level features in the framework of a high-level concept using a knowledge management. Starting from a representation of the three most frequently used features i.e. color, texture and shape, cooperation techniques are proposed in order to use the expert knowledge in the combination process. Two types of cooperation have been defined; the first is without categorization and the second with categorization. The second approach allows to make a selection of the query category in order to simplify queries in large image databases. Finally, in order to avoid the standard relevance feedback stage, a competition technique has been proposed allowing an unsupervised refinement of the submitted query. The proposed techniques have shown their performances and their robustness and are generalizable to all types of image databases.