Soutenance de Thèse – Matthieu MULLER

Thèse préparée au sein du laboratoire Gipsa-lab et de la Faculté de génie électrique et informatique de l’Université d’Islande.

Abstract :

Image processing is essential for extracting meaningful information from sensor-acquired visual data, with applications ranging from medical imaging to remote sensing. Raw images often suffer from noise, blur, limited resolution, and CFA artifacts due to sensor and environmental constraints, making restoration a critical step. Traditional model-based approaches, grounded in physical and mathematical principles, provide interpretability but often fail to capture the complexity of real-world degradations.
Deep learning has recently revolutionized image restoration by learning powerful mappings from degraded to clean images, though it requires large labeled datasets, significant computational resources, and offers limited interpretability. Hybrid approaches that combine model-based
priors with the flexibility of neural networks aim to overcome these limitations by improving robustness, reducing data demands, and providing more explainable solutions. This thesis explores such hybrid methods, with a particular focus on the universal demosaicking problem, that is reconstructing full-color images from gray-scale CFA inputs.

Membres du Jury :

Cédric RICHARD, University Côte d’Azur, Président
Jon Yngve HARDEBERG, Norwegian University of Science and Technologie, Rapporteur
Paolo GAMBA, University of Pavia, Rapporteur
Abdelaziz KALLEL, Digital Research Center of Sfax, Examinateur
Yusuke MONNO, Institute of Science Tokyo, Examinateur