A Discussion of Nonlinear Variants of Biased Discriminants for Interactive Image Retrieval

Abstract

During an interactive image retrieval process with relevance feedback, kernel-based or boosted learning algorithms can provide superior nonlinear modeling capability. In this paper, we discuss such nonlinear extensions for biased discriminants, or BiasMap [1, 2]. Kernel partial alignment is proposed as the criterion for kernel selection. The associated analysis also provides a gauge on relative class scatters, which can guide an asymmetric learner, such as BiasMap, toward better class modeling. We also propose two boosted versions of BiasMap. Unlike existing approach that boosts feature components or vectors to form a composite classifier, our scheme boosts linear BiasMap toward a nonlinear ranker which is more suited for small-sample learning during interactive image retrieval. Experiments on heterogeneous image database retrieval as well as small sample face retrieval are used for performance evaluations.