Learning of Filter-Based Regularizers
Available
Master Semester Project
Project: 00448
In this project, you will train regularizers for solving medical image reconstruction problems. The starting point is a simple and interpretable regularizer architecture that leads to very competitive performance when compared with deep neural network approaches. Due to its specific design, the approach closely resembles traditional signal and image processing models. For the training, it is necessary to efficiently solve the underlying optimization problem, where several approaches are possible. Throughout this project, you will explore these possibilities and use recent PyTorch libraries to make the training and evaluation more efficient. As part of this, you will also review the theory for the deployed optimization methods.
- Supervisors
- Sebastian Neumayer, sebastian.neumayer@epfl.ch
- Stanislas Ducotterd, stanislas.ducotterd@epfl.ch