Learning Taxonomies in Large Image Databases

Growing image collections have created a need for effective retrieval mechanisms. Although content-based image retrieval systems have made huge strides in the last decade, they often are not sufficient by themselves. Many databases, such as those at Flickr are augmented by keywords supplied by its users. A big stumbling block however lies in the fact that many keywords are actually similar or occur in common combinations which is not captured by the linear metadata system employed in the databases. This paper proposes a novel algorithm to learn a visual taxonomy for an image database, given only a set of labels and a set of extracted feature vectors for each image. The taxonomy tree could be used to enhance the user search experience in several ways. Encouraging results are reported with experiments performed on a subset of the well known Corel Database.