Biased minimax probability machine active learning for relevance feedback in content-based image retrieval

Abstract

In this paper we apply Biased Minimax Probability Machine (BMPM) to address the problem of relevance feedback in Content-based Image Retrieval (CBIR). In our proposed methodology we treat relevance feedback task in CBIR as an imbalanced learning task which is more reasonable than traditional methods since the negative instances largely outnumber the positive instances. Furthermore we incorporate active learning in order to improve the framework performance, i.e., try to reduce the number of iterations used to achieve the optimal boundary between relevant and irrelevant images. Different from previous works, this model builds up a biased classifier and achieves the optimal boundary using fewer iterations. Experiments are performed to evaluate the efficiency of our method, and promising experimental results are obtained. © Springer-Verlag Berlin Heidelberg 2006.