Colour Image Segmentation using Texems

We present two methods to perform colour image segmentation using a generative three-dimensional model that is based on the assumption that an image can be generated through an overlapped placement of a few primitive, exemplar image patches, i.e. texems. Multiscale analysis is used in order to capture sufficient image features and pixel neighbourhood interactions at relatively lower computational costs. Experimental results on synthetic and real images are presented to demonstrate this as a promising alternative approach to popular discriminative methods.