Dissertation in Computer Science with Samuel Willingham

About the thesis

When capturing or processing images, one often encounters perturbations such as noise or blur. Recovering a clean image from such a degraded observation is called an inverse problem. Unfortunately, many inverse problems are ill-posed and do not permit a straightforward solution, leading to a need for additional constraints on what constitutes a clean image.

This additional information is called the prior and enters image restoration in the form of regularization, commonly performed by deep neural networks. This work aims to improve understanding of image regularization by investigating theory and training approaches for regularizers. The first two works investigate the training of deep priors in Maximum a Posteriori (MAP) algorithms; this is done by extending deep equilibrium (DEQ) models, which allow the training of full iterative frameworks, to the multi-task training of MAP algorithms.

The third and fourth works investigate an application of MAP training to generative restoration including a proposed loss function for MAP training. Results showed DEQ multi task training in combination with mathematically sound losses can lead to strong reconstruction performance without affecting inference time.

More information

Date: 30 November 2026

Time:14:00

Location: Inria Centre, Rennes University and via Zoom (the zoom link will be provided closer to the public defence)

Doctoral thesis: "Learning Methods for Inverse Problems in Imaging".

Defendant: Samuel Willingham. 

Supervisors and chairs: Professor Mårten Sjöström, Mid Sweden University and Dr. Christine Guillemot, Rennes University.

Review Committee:  

Dr. Andrés Almansa, MAP5 Université Paris Cité, FR 

Prof. Saïd Moussaoui, Centrale de Nantes, FR 

Dr. Luca Calatroni, University of Genoa, IT 

Prof. Gabriele Facciolo, Centre Borelli, ENS Paris Saclay, FR

Dr. Charles Kervrann, Inria Rennes, FR