An optimal transport analogue of the Rudin Osher Fatemi model and its corresponding multiscale theory

Abstract

We develop a theory for image restoration with a learned regularizer that is analogous to that of Meyer's characterization of solutions of the classical variational method of Rudin-Osher-Fatemi (ROF). The learned regularizer we use is a Kantorovich potential for an optimal transport problem of mapping a distribution of noisy images onto clean ones, as first proposed by Lunz, \"Oktem and Sch\"onlieb. We show that the effect of their restoration method on the distribution of the images is an explicit Euler discretization of a gradient flow on probability space, while our variational problem, dubbed Wasserstein ROF (WROF), is the corresponding implicit discretization. We obtain our geometric characterisation of the solution in this setting by first proving a more general convex analysis theorem for variational problems with solutions characterised by projections. We then use optimal transport arguments to obtain our WROF theorem from this general result, as well as a decomposition of a transport map into large scale "features" and small scale "details", where scale refers to the magnitude of the transport distance. Further, we leverage our theory to analyze two algorithms which iterate WROF. We refer to these as iterative regularization and multiscale transport. For the former we prove convergence to the clean data. For the latter we produce successive approximations to the target distribution that match it up to finer and finer scales. These algorithms are in complete analogy to well-known effective methods based on ROF for iterative denoising, respectively hierarchical image decomposition. We also obtain an analogue of the Tadmor Nezzar Vese energy identity which decomposes the Wasserstein 2 distance between two measures into a sum of non-negative terms that correspond to transport costs at different scales.