Bayesian Active Learning in Relevance Feedback for Image Retrieval
Jun Wu, Yingling Fu, Mingyu Lu · 2008
One of the fundamental problems in Content-Based Image Retrieval (CBIR) has been the gap between low-level visual features and high-level semantic concepts. To narrow the gap, relevance feedback (RF) is introduced into CBIR. However, most RF methods are challenged by small size sample collection and asymmetric sample distributions between the positive and the negative samples. In this paper, a Bayesian active learning (BAL) mechanism is proposed to overcome these problems. First, by defining the confidence, we design a new selection criterion for the images with the low confidence, which to be labeled by user, and then the learner is retrained by the most informative samples obtained from the last round feedback. Moreover, different learning strategies are used for estimating the distributions of the positive and the negative samples. Based on above methods, the retrieval performance can be enhanced. The experimental results on Corel image database demonstrate the effectiveness of the proposed algorithm.