Relevance feedback based on parameter estimation of target distribution

Abstract

Relevance feedback formulations have been proposed to refine query result in content-based image retrieval in the past few years. Many of them focus on a learning approach to solve the feedback problem. In this paper, we present an expectation maximization approach to estimate the user's target distribution through user's feedback. Furthermore, we describe how to use the maximum entropy display to fully utilize user's feedback information. We detail the process and also demonstrate the result through experiments.