Image annotation research has demonstrated success on test data for focused domains. Unfortunately, extending these techniques to the broader topics found in real world data often results in poor performance. This paper proposes a novel approach that leverages WordNet and ImageNet capabilities to annotate images based on local text and image features. Signatures generated from ImageNet images based on WordNet synonymous sets are compared using Earth Mover's Distance against the query image and used to rank order surrounding words by relevancy. The results demonstrate effective image annotation, producing higher accuracy and improved specificity over the ALIPR image annotation system. Abstract © AAAI.
26th International Florida Artificial Intelligence Research Society Conference, FLAIRS 2013
Noel, G. E., & Peterson, G. L. (2013). Context-driven image annotation using ImageNet. 26th International Florida Artificial Intelligence Research Society Conference, FLAIRS 2013, 462–467.