SUPR
Learning methods for inverse problems on novel imaging modalities
Dnr:

NAISS 2024/22-864

Type:

NAISS Small Compute

Principal Investigator:

Samuel Willingham

Affiliation:

Mittuniversitetet

Start Date:

2024-06-14

End Date:

2025-07-01

Primary Classification:

10206: Computer Engineering

Webpage:

Allocation

Abstract

Inverse problems appear all across signal processing and involve the reconstruction of a degraded signal to match a clean ground truth signal. In imaging, this pertains to tasks like super-resolution (digital zoom), in-painting, deblurring, or denoising. For higher dimensional images, like light field images, this includes angular super-resolution, meaning the synthesis of new viewing angles from a set of reference views. Given that these inverse problems are often ill-posed, regularization is crucial. Deep learning excels for this and many neural network-based approaches are currently being investigated as regularizers in iterative procedures. This work leverages deep equilibrium models to train deep regularizers in iterative procedures via the corresponding fixed-point. The overall aim is to train neural networks that allow for high-quality image restoration and generation in the context of inverse problems.