Learning to Recommend Tags for On-line Photos

Recommending text tags for on-line photos is useful for on-line photo services. We propose a novel approach to tag recommendation by utilizing both the underlying semantic correlation between visual contents and text tags and the tag popularity learnt from realistic on-line photos. We apply our approach to a database of real on-line photos and evaluate its performance by both objective and subjective evaluation. Experiwith ments demonstrate the improved performance of the proposed approach compared the state-of-the-art techniques in the literature.