Automatic annotation of weakly-tagged social images on flickr using latent topic discovery of multiple groups

  • Authors:

  • Ye Chen;Jian Shao;Ke Zhu

  • Affiliations:

  • College of Computer Science and Technology, Hang Zhou, China;College of Computer Science and Technology, Hang Zhou, China;College of Computer Science and Technology, Hang Zhou, China

  • Venue:
  • AMC '09 Proceedings of the 2009 workshop on Ambient media computing
  • Year:
  • 2009

Quantified Score

Hi-index 0.01

Visualization

Abstract

At photo-sharing websites like Flickr, a lot of "weakly-tagged" social images can not be effectively retrieved because they are noisily or sparsely tagged. Since users in Flickr often recommend their uploaded images to multiple associated groups according to the latent topics in each image, we propose a novel two-stage approach to automatically annotate these weakly-tagged social images. In the first stage called as topic-guided tag refinement, the latent topics in each group are beforehand discovered by the Latent Dirichlet Allocation model, then those noisy tags are filtered in group level and topic-relevant tags are re-ranked in image level before and after tag prorogation among similar images respectively. In the second stage called as hierarchical multi-group tag fusion, the hierarchical topic structure among multiple groups is beforehand discovered by WordNet, and is then used to fuse the generated tags from multiple groups in hierarchical way. Experiments on 93481 images from three Flickr groups show the effectiveness of our proposed approach.