Robust Nonparametric Relevance Feedback for Image Retrieval

Most interactive, “query-by-example” based image retrieval systems employ relevance feedback technique for bridging the gap between the user-defined high-level concept and the low-level image representation in the feature space. We propose in this paper a unified relevance feedback methodology that offers flexibility in capturing user perception and at the same time robustness to deal with limited training images. A generalized additive model based nonparametric probabilistic approach is adopted for flexibility. A generalized ellipsoid based parametric model with outlier rejection is proposed for robustness. Our approach initially assumes a unimodal user perception, and depending on the size of the outliers infers a multi-modal perception, and switches to nonparametric mode. Experimental results with simulated training set are presented to demonstrate the validity and effectiveness of the proposed relevance feedback technique. We also report results on real image databases and show the effect of our algorithm on the end-to-end retrieval performance.