Deep Denoising: An Overview

by Luis Felipe Barba Flores (PSI - Paul Scherrer Institut)

Thursday 29 Oct 2026, 11:30 → 13:00 Europe/Zurich

OFLG/402 or online

Abstract:
Recovering clean signals from noisy observations is a foundational problem in scientific imaging, and deep neural networks have substantially redefined what is achievable. This talk surveys the landscape of deep denoising methods, organized around a central question: what supervision is available at training time? We begin with the fully supervised setting, where paired clean targets exist, and trace the progression toward regimes where clean data is scarce or entirely absent, covering methods that exploit pairs of independent noisy realizations, techniques that learn from a single corrupted image by leveraging statistical independence assumptions across pixels or measurements, and approaches grounded in score estimation and Tweedie's formula that connect denoising to the underlying data distribution. We also examine how the structure of the measurement process itself, particularly in tomographic and other linear inverse problems, can be turned into a self-supervision signal, removing the need for any clean reference. We might also discuss diffusion-based posterior samplers, which shift the goal from a single point estimate to characterizing the full posterior over clean images given noisy observations. Throughout, we highlight the tradeoffs between data requirements, noise model assumptions, and the nature of the outputs produced. Finally, I will touch upon some applications in computed tomography and how this denoising techniques enable better data compression. 

for online participation: TEAMS link here