Elhadji Cisse Faye: Bayesian inversion with deep learning-driven priors.

par M. Elhadji Cisse Faye (IDP-Orléans)

jeudi 9 oct. 2025, 10:30 → 11:30 Europe/Paris

Salle de séminaire (IDP-Orléans)

This thesis addresses the restoration of degraded images, a key challenge in medical imaging, astrophysics, and microscopy. The goal is to recover an image from noisy observations while quantifying uncertainty. We adopt a Bayesian approach that combines explicit probabilistic modeling with deep neural networks trained for denoising. We design novel sampling algorithms, including a Gibbs-type scheme that integrates Langevin dynamics adapted to the geometry of the problem. Our methods achieve high-quality reconstructions and provide rigorous uncertainty estimates, enabling more interpretable and reliable image analysis.