Asymmetric Bayesian Learning for Image Retrieval with Relevance Feedback

Bayesian learning (BL) based relevance feedback (RF) schemes plays a key role for boosting image retrieval performance. However, traditional BL based RF schemes are often challenged by the small example problem and asymmetrical training example problem. This paper presents a novel scheme that embeds the query point movement (QPM) technique into the Bayesian framework for improving RF performance. In particular, we use an asymmetric learning methodology to determine the parameters of Bayesian learner, thus termed as asymmetric Bayesian learning. For one thing, QPM is applied to estimate the distribution of the relevant class by exploiting labeled positive and negative examples. For another, a semi-supervised learning mechanism is used to tackle the scarcity of negative examples. Concretely, a random subset of the unlabeled images is selected as the candidate negative examples, of which the problematic data are then eliminated by using QPM. Then, the cleaned unlabeled images are regarded as additional negative examples which are helpful to estimate the distribution of the irrelevant class. Experimental results show that the proposed scheme is more effective than some existing approaches.