In this work we deal with the problem of summarizing image collections that correspond to a single event each. For this, we adopt a clustering-based approach, and we perform a comparative study of different clustering algorithms and image representations. As part of this study, we propose and examine the possibility of using trained concept detectors so as to represent each image with a vector of concept detector responses, which is then used as input to the clustering algorithms. A technique which indicates which concepts are the most informative ones for clustering is also introduced, allowing us to prune the employed concept detectors. Following the clustering, a summary of the collection (thus, also of the event) can be formed by selecting one or more images per cluster, according to different possible criteria. The combination of clustering and concept-based image representation is experimentally shown to result in the formation of clusters and summaries that match well the human expectations.