Path Variation and Image Segmentation

Pablo Andrés Arbeláez, Laurent D. Cohen

Abstract: We study a notion of variation for real valued two variable functions called the path variation and we discuss its application as a low-level image segmentation method. For this purpose, we characterize the path variation as an energy in the framework of minimal paths. In this context, the definition of an energy and the selection of a set of source points determine a partition of the image domain. The problem of choosing a relevant set of sources is addressed through a nonlinear diffusion filtering.