Automated Annotation of Landmark Images Using Community Contributed Datasets and Web Resources

A novel solution to the challenge of automatic image annotation is described. Given an image with GPS data of its location of capture, our system returns a semantically-rich annotation comprising tags which both identify the landmark in the image, and provide an interesting fact about it, e.g. “A view of the Eiffel Tower, which was built in 1889 for an international exhibition in Paris”. This exploits visual and textual web mining in combination with content-based image analysis and natural language processing. In the first stage, an input image is matched to a set of community contributed images (with keyword tags) on the basis of its GPS information and image classification techniques. The depicted landmark is inferred from the keyword tags for the matched set. The system then takes advantage of the information written about landmarks available on the web at large to extract a fact about the landmark in the image. We report component evaluation results from an implementation of our solution on a mobile device. Image localisation and matching offers 93.6% classification accuracy; the selection of appropriate tags for use in annotation performs well (F1M of 0.59), and it subsequently automatically identifies a correct toponym for use in captioning and fact extraction in 69.0% of the tested cases; finally the fact extraction returns an interesting caption in 78% of cases.