Despite the major effort put into the creation of Content-Based Image Retrieval (CBIR) systems during the last decade, the solutions available are still not satisfying for generic purposes. The most severe issue seems to be the so-called "semantic gap". It is feasible to define and use domain specific feature vectors on a low level and use this information for a similarity based retrieval. Yet, mapping these to higher level semantics remains difficult. This research investigates a domain-independent way of automatized image categorization. A CBIR query language is constructed to build query-like descriptors for each category to be learned. The proposed learning algorithm is based on decision-trees. The resulting descriptors are aimed to be understandable and modifiable by expert users. A case-study is presented to support these claims.