Today's systems for Content-Based Image Retrieval (CBIR) suffer from several drawbacks: First, user interfaces are much too complicated for average users. Second, the quality of results tends to be low. Finally, querying performance with often long reply times is unsatisfactory. We have developed a model for CBIR searches intended to overcome these drawbacks where the user has to select only one or more example images to initiate a query. Out of the examples the actual query including feature selection and weighting is generated automatically. The results of the first query may later on be refined by relevance feedback. We discuss the major components necessary for our approach and the results achieved in our test environment.
Proceedings of Digital Libraries Conference, Tokyo, Japan, 2000