Bayesian Imaging in the Low-Photon Regime: Data-Driven Priors and Algorithms

Title

Bayesian Imaging in the Low-Photon Regime: Data-Driven Priors and Algorithms

Speaker

Teresa Klatzer - University of Edinburgh

Speaker

Teresa Klatzer, University of Edinburgh

Abstract

Reconstructing images in low-photon scenarios, such as astronomy, biology, and medical imaging, poses severe challenges due to high uncertainty and non-Gaussian noise. This talk explores Bayesian computational approaches that integrate data-driven priors with advanced posterior sampling methods to achieve robust reconstruction and principled uncertainty quantification. I will outline how imaging problems can be formulated within a Bayesian framework and highlight Langevin-based methodologies particularly suited to these tasks. We focus on recent developments for low-photon Poisson imaging using plug-and-play (PnP) Langevin sampling. Standard PnP Langevin algorithms are not well-suited for Poisson data due to high uncertainty, exploding gradients, and non-negativity constraints. To address these issues, we propose two approaches: (i) an accelerated PnP Langevin method with boundary reflections and a likelihood approximation, and (ii) a mirror sampling algorithm utilizing Riemannian geometry to handle constraints and poor likelihood regularity directly. The methods are evaluated on tasks such as image deconvolution and low-dose computer tomography. Empirically, these sampling-based approaches yield improved reconstructions and informative uncertainty estimates, offering a promising foundation for future advances in low-photon imaging.

Bio

Teresa Klatzer is a Postdoc at the University of Edinburgh, Scotland, working with Konstantinos Zygalakis funded by the Prob_AI Doctoral Prize Fellowship. Her research interests lie at the intersection of Bayesian computation, machine learning and low-photon imaging problems. She recently completed her PhD at Edinburgh, with a thesis entitled “Bayesian Imaging with data-driven priors”. Her work includes novel methodologies for plug-and-play Langevin sampling, Poisson inverse problems and accelerated sampling for imaging.

When

Thursday, October 30th, h 12:00

Where

Room 322, DIBRIS Via Dodecaneso 35