Automatic Generation of Semantic Fields for Annotating Web Images

The overwhelming amounts of multimedia contents have triggered the need for automatically detecting the semantic concepts within the media contents. With the development of photo sharing websites such as Flickr, we are able to obtain millions of images with user-supplied tags. However, user tags tend to be noisy, ambiguous and incomplete. In order to improve the quality of tags to annotate web images, we propose an approach to build Semantic Fields for annotating the web images. The main idea is that the images are more likely to be relevant to a given concept, if several tags to the image belong to the same Semantic Field as the target concept. Semantic Fields are determined by a set of highly semantically associated terms with high tag co-occurrences in the image corpus and in different corpora and lexica such as WordNet and Wikipedia. We conduct experiments on the NUS-WIDE web image corpus and demonstrate superior performance on image annotation as compared to the state-of-the-art approaches.