Learning in content-based image retrieval

摘要

In this paper we address several aspects of the learning problem in content-based image retrieval (CBIR). First, we introduce the linear and kernel-based biased discriminant analysis, or BiasMap, to fit the unique nature of relevance feedback as a small sample biased classification problem. Secondly, a WARF (word association via relevance feedback) formula is presented for learning keyword relations during the process of relevance feedback. We also introduce our new user interface for CBIR, ImageGrouper, which is designed to support more sophisticated user feedbacks and annotations. Finally, we use the D-EM (Discriminant-EM) algorithm as a way of exploiting unlabeled data in CBIR and offer some insights as to when unlabeled data will help.