Towards More Precise Social Image-Tag Alignment

Large-scale user contributed images with tags are increasingly available on the Internet. However, the uncertainty of the relatedness between the images and the tags prohibit them from being precisely accessible to the public and being leveraged for computer vision tasks. In this paper, a novel algorithm is proposed to better align the images with the social tags. First, image clustering is performed to group the images into a set of image clusters based on their visual similarity contexts. By clustering images into different groups, the uncertainty of the relatedness between images and tags can be significantly reduced. Second, random walk is adopted to re-rank the tags based on a cross-modal tag correlation network which harnesses both image visual similarity contexts and tag co-occurrences. We have evaluated the proposed algorithm on a large-scale Flickr data set and achieved very positive results.