Image annotation using semantic clusters from Flickr

Image annotation is an important topic for effective image retrieval. In this paper, we propose an image annotation method using semantic clusters from Flickr, aiming to find more semantically similar images with the query image. For an image to be annotated, an initial word is firstly provided to search the corresponding semantic clusters on Flickr. The representative images for the clusters are also crawled. Then a content based method is utilized to decide the proper semantic cluster the query image belongs to. Finally a set of candidate annotations are extracted and ranked from top N visual similar images in the specific semantic cluster. The proposed algorithm avoids the problem of word ambiguity in existing methods, and provides more semantically consistent annotations. Experimental results show that the performance of image annotation could be improved when incorporating semantic clusters.