Theoretical Foundations for Discrete Forward-and-Backward Diffusion Filtering

Abstract. Forward-and-backward (FAB) diffusion is a method for sharpening blurry images (Gilboa et al. 2002). It combines forward diffusion with a posi-tive diffusivity and backward diffusion where negative diffusivities are used. The well-posedness properties of FAB diffusion are unknown, and it has been ob-served that standard discretisations can violate a maximum-minimum principle. We show that for a novel nonstandard space discretisation which pays specific attention to image extrema, one can apply a modification of the space-discrete well-posedness and scale-space framework of Weickert (1998). This allows to establish well-posedness and a maximum-minimum principle for the resulting dynamical system. In the fully discrete 1-D case with an explicit time discretisa-tion, a maximum-minimum principle and total variation reduction are proved in spite of the fact that negative diffusivities may appear. This provides a theoretical justification for applying FAB diffusion to digital images.