Automatic refinement of keyword annotations for web image search

Automatic image annotation is fundamental for effective image browsing and search. With the increasing size of image collections such as web images, it is infeasible to manually label large numbers of images. Meanwhile, the textual information contained in the hosting web pages can be used as approximate image description. However, such information is not accurate enough. In this paper, we propose a framework to utilize the visual content, the textual context, and the semantic relations between keywords to refine the image annotation. The hypergraph is used to model the textual information and the semantic relation is deduced by WordNet. Experiments on large-scale dataset demonstrate the effectiveness of the proposed method.