Nonparametric discriminant analysis in relevance feedback for content-based image retrieval

Relevance feedback (RF) has been widely used to improve the performance of content-based image retrieval (CBIR). How to select a subset of features from a large-scale feature pool and to construct a suitable dissimilarity measure are key steps in RF. Biased discriminant analysis (BDA) has been proposed to select features during relevance feedback iterations. However, BDA assumes all positive feedbacks form a single Gaussian distribution, which may not be the case for CBIR. Although kernel BDA can overcome the drawback to some extent, the kernel parameter tuning makes the online learning unfeasible. To avoid the parameter tuning problem and the single Gaussian distribution assumption in BDA, we construct a new nonparametric discriminant analysis (NDA). To address the small sample size problem in NDA, we introduce the regularization method and the -space method. Because the regularization method may meet the ill-posed problem and the -space method will lose some discriminant information, we proposed here a full-space method. The proposed full-space NDA is demonstrated to outperform BDA based RF significantly based on a large number of experiments in Corel database with 17,800 images.