Evolving Content-Driven Superpixels for Accurate Image Representation

A novel approach to superpixel generation is presented that aims to reconcile image information with superpixel coverage. It is described as content-driven as the number of superpixels in any given area is dictated by the underlying image properties. By using a combination of well-established computer vision techniques, superpixels are grown and subsequently divided on detecting simple image variation. It is designed to have no direct control over the number of superpixels as this can lead to errors. The algorithm is subject to performance metrics on the Berkeley Segmentation Dataset including: explained variation; mode label analysis, as well as a measure of oversegmentation. The results show that this new algorithm can reduce the superpixel oversegmentation and retain comparable performance in all other metrics. The algorithm is shown to be stable with respect to initialisation, with little variation across performance metrics on a set of random initialisations.