Unsupervised image segmentation remains a challenging problem in computer vision. Among the several techniques, graph-based methods are becoming increasingly popular. Particularly, a graph-theoretic concept known as dominant set has been shown to be useful in a variety of applications. However, segmenting large images with this method is intractable due to its scaling behavior with the number of pixels. This paper introduces an efficient method for image segmentation using dominant sets. This is accomplished by building the graph over superpixels rather than pixels, which lowers the number of nodes that need to be processed. Based on experimental results, we have achieved higher accuracy and faster processing time than previous methods.