The identification of categories in image databases usually relies on clustering algorithms that only exploit the feature-based similarities between images. The addition of semantic information should help improve the results of the categorisation process. Pairwise constraints between some images are easy to provide, even when the user has a very incomplete prior knowledge of the image categories that one can expect to find in a database. A categorisation approach relying on such semantic information is called semi-supervised clustering. A new semi-supervised clustering algorithm, pairwise-constrained competitive agglomeration, is presented on the basis of a fuzzy cost function that takes pairwise constraints into account. Evaluations show that with a rather low number of constraints this algorithm can significantly improve the categorisation.