6533b854fe1ef96bd12ade36

RESEARCH PRODUCT

Interactive Image Retrieval Using Smoothed Nearest Neighbor Estimates

subject

description

Relevance feedback has been adopted by most recent Content Based Image Retrieval systems to reduce the semantic gap that exists between the subjective similarity among images and the similarity measures computed in a given feature space. Distance-based relevance feedback using nearest neighbors has been recently presented as a good tradeoff between simplicity and performance. In this paper, we analyse some shortages of this technique and propose alternatives that help improving the efficiency of the method in terms of the retrieval precision achieved. The resulting method has been evaluated on several repositories which use different feature sets. The results have been compared to those obtained by the nearest neighbor approach in its standard form, suggesting a better performance.