Improving image reconstruction algorithms with model mismatches

par Pierre Weiss (IMT)

vendredi 12 mai 2023, 11:00 → 12:00 Europe/Paris

Salle K. Johnson (IMT)

In this talk, I will describe an overview of our recent attempts to improve imaging systems using deep learning techniques. While efficient tools are now available when the forward model describing the acquisition device is well calibrated, they usually perform terribly when model mismatches occur. In that case, we will see that specific neural network architectures make it possible to identify the system from the measurements themselves. This yields self-calibrated reconstruction algorithms. Application examples include optical imaging, MRI or computerized tomography. 
 

 

This talk will also be an opportunity to discuss potential interactions with the CBI (Centre de Biologie Intégrative), where I just created a group. Many biologists produce complex data sets and are expressing a strong will to discuss modelling issues.