Pablo ARIAS : Habilitation à diriger des recherches

Pablo ARIAS

Video restoration: from patch-based methods to self-supervised learning

Abstract

I will present contributions on image and video restoration problems, such as denoising and super-resolution. In the first part I will describe Bayesian approaches based on probabilistic models for image patches applied to video denoising. In the second part I will propose approaches based on deep learning for multi-image restoration problems found in practical applications. In these cases, the standard supervised training is not feasible due to the lack of real data with known ground truth. I will present a framework for self-supervised learning for multi-image problems that allows training machine learning models directly over the real degraded data, without requiring ground truth reconstructions.

Jury