Implementing Relevance Feedback Techniques for Large Image Collections Efficiently Klemens Böhm, Ana Stojanovic, and Roger Weber   View Paper (PDF)   Return to Papers Abstract Relevance Feedback is a powerful technique to raise the quality of search results in image databases. A number of relevance feedback models have been proposed. Frequently, these solutions apparently do not scale with the number of images. Another issue is that some relevance feedback models support only simple similarity queries, but not similarity queries with a more complex structure, e.g., more than one reference image. This article in turn describes the design and implementation of a relevance feedback engine that supports relevance feedback in its full generality. To this end, we have developed a query language for similarity search with the following characteristics: it is expressive enough to map user feedback to a statement in the language for any of the relevance feedback models under consideration here. At the same time, it is not excessively complex and allows for an extremely efficient implementation, even for large data sets. DiSC'02 © 2003 Association for Computing Machinery