Robust image annotation refinement via graph-based learning

Image annotation has been an active research topic in recent years. However, the state of art image annotation methods are often unsatisfactory, in this paper, we presented a novel image annotation refinement to improve the performance of automatic image annotation. Firstly, the initial pair-wise similarities of words is computed based on the co-occurrence of training sets, Then the topic relation is mined by generating the topic bag. Finally, the candidate annotations are re-ranked by embedding the refined word relation. The experiments over Corel images have shown that embedding topic relation is beneficial in image annotation.