Dictionary based image segmentation

Abstract

We propose a method for weakly supervised segmentation of natural images, which may contain both textured or non-textured regions. Our texture representation is based on a dictionary of image patches. To divide an image into separated regions with similar texture we use an implicit level sets representation of the curve, which makes our method topologically adaptive. In addition, we suggest a multi-label version of the method. Finally, we improve upon a similar texture representation, by formulating the computation of a texture probability in terms of a matrix multiplication. This results in an efficient implementation of our segmentation method. We experimentally validated our approach on a number of natural as well as composed images.