This paper describes a novel approach that automatically refines the image annotations generated by a non-parametric density estimation model. We re-rank these initial annotations following a heuristic algorithm, which uses semantic relatedness measures based on keyword correlation on the Web. Existing approaches that rely on keyword co-occurrence can exhibit limitations, as their performance depend on the quality and coverage provided by the training data. Additionally, WordNet based correlation approaches are not able to cope with words that are not in the thesaurus. We illustrate the effectiveness of our Web-based approach by showing some promising results obtained on two datasets, Corel 5k, and ImageCLEF2009.