Investigating Orthogonal Convolutions for Denoising
Spring 2023
Master Semester Project
Project: 00440

Training convolutional neural networks with a Lipschitz constraint allows them to be inserted in iterative algorithms to solve inverse problems with the guarantee that the algorithm converges. The goal of this project is to get familiar with different parametrizations of orthogonal convolutional layers and to compare their performance by training Lipschitz-constrained CNN denoisers. This project requires good PyTorch knowledge and a good understanding of the theory from the Signals & Systems course which is extended to vector-valued signals and matrix-valued filters in the theory of orthogonal convolutions.
- Supervisors
- Stanislas Ducotterd, stanislas.ducotterd@epfl.ch
- Sebastian Neumayer, sebastian.neumayer@epfl.ch
- Alexis Goujon, alexis.goujon@epfl.ch