MAP-based image tag recommendation using a visual folksonomy

Image Tag Completion via Image-Specific and Tag-Specific Linear Sparse Reconstructions

“…JEC [10] and TagProp [3]), tag recommendation approaches (i.e. Vote+ [11] and Folksonomy [5]) and recently proposed unified tag refinement frameworks of denoising and completion (i.e. LR [16] and SUG [9]) as baselines for comparison.…”

“…M. Guillaumin et al [3] put forward another auto-annotation model named TagProp, which adopts discriminative metric learning methods in nearest neighbour models and maintains the state-of-the-art performance. Tag recommendation [11,14,5] is a trade-off between autoannotation and manual tagging, which is to recommend semantically related tags to a user while he is annotating an image online. B. Sigurbjörnsson and R. v. Zwol [11] introduced a generic tag recommendation method that deploys the collective knowledge residing in images.…”

“…B. Sigurbjörnsson and R. v. Zwol [11] introduced a generic tag recommendation method that deploys the collective knowledge residing in images. And S. Lee et al [5] formulated tag recommendation as a maximum a posteriori (MAP) problem using a visual folksonomy.…”

“…The integrated objective function can be demonstrated to be convex, meaning that there exists a global optimal solution. Then the optimal and can be merged for obtaining a reconstructed tagging vector ′ for the target image, as shown in formula (5).…”

Tag refinement in an image folksonomy using visual similarity and tag co-occurrence statistics

“…As such, we modify the way the conditional probability PðijtÞ is estimated in order to mitigate the influence of tag noise in an image folksonomy. In particular, instead of computing the conditional probability PðijtÞ over the whole image folksonomy F, we propose to estimate the conditional probability PðijtÞ over a newly constructed folksonomy Q [47]. The folksonomy Q, which is derived from the original folksonomy F, contains images that are visually similar to the seed image i.…”