Meta-tag propagation by co-training an ensemble classifier for improving image search relevance

The ever-increasing gigantic amount of images over the Web necessitates automatic schemes for meta-tagging content descriptions such as object categories. These meta-tags are essential to text-based image search engines to improve their search relevance. Traditional supervised scheme is not suitable for this task because it needs too much manual labelling efforts and yet is hard to scale to a large number of object categories. Notice that in the search scenarios, the meta-tagging does not need to be perfect to help improve relevance because the available text tags and user click-through logs can partially rectify the inaccurate information. A weakly supervised scheme would be ideal when only sporadic labelled examples are exploited in spite of the expected loss in tagging accuracy. In this paper, we develop a novel weakly semi-supervised ensemble classifier trained based on a co-training framework for this tagging task. In essence the meta-tags are recursively propagated from the sparsely tagged examples to the un-tagged ones. Preliminary experiments on benchmark database such as Graz02 demonstrate the efficacy of the proposed approach.