Effective automatic image annotation via integrated discriminative and generative models

In this paper, we present a novel image annotation method that leverages on the advantages of both generative and discriminative models. To label an image, we first identify a visual neighborhood in the training image set based on generative approach. Then, the neighborhood is refined by an optimal discriminative hyperplane tree classifier based on concept feature. The tree classifier is built according to a local topic hierarchy, which is adaptively constructed by exploiting the semantic contextual correlations of the corresponding visual neighborhood. Experiments conducted on the ECCV2002 and TRECVID 2005 benchmarks demonstrate the effectiveness and efficiency of the proposed method.