Skip to Main Content
Statistical generative model based image annotation propagates the semantic labels of the training images to the unlabeled ones according to their visual generative probabilities. However, it suffers from the problem of "semantic gap", that is, sometimes visual similarity does not reflect semantic similarity. In order to alleviate this problem, we propose a novel image annotation approach which combines the advantages of the generative model and discriminative classification. Based on generative model, we exploit the local discriminants of the visual similar training images (neighborhood) of the unlabeled image. The semantic similar images in the neighborhood are grouped as topics by singular value decomposition (SVD). The discriminative information between different topics is exploited to obtain the semantic relevant topic, which reduces the influence of the images with high visual similarity but irrelevant semantics. Thus, the joint probability of the semantic keyword and the unlabeled image estimated on the obtained relevant topic is more accurate. The experimental results on the ECCV2002 benchmark (P. Duygulu et al., 2002) show that our method outperforms state-of-the-art annotation models MBRM and ASVM-MIL.