Title: Learning to mean-shift in $O(1)$ for Bayesian image restoration
Authors: Siavash Bigdeli - EPFL (Switzerland) [presenting]
Meiguang Jin - University of Bern (Switzerland)
Paolo Favaro - University of Bern (Switzerland)
Matthias Zwicker - University of Maryland (United States)
Abstract: Finding strong oracle priors is an important topic for solving ill-posed problems. We show how denoising autoencoders (DAEs) learn to mean-shift in $O(1)$, and how we leverage this to employ DAEs as generic priors for the task of image restoration. We also discuss the case of Gaussian DAEs in a Bayesian framework, where the degradation parameters (e.g. noise and/or blur kernel) are unknown. Experimental results demonstrate state-of-the-art performance of the proposed DAE priors in image deblurring and super-resolution.