Apprentissage de méthodes itératives pour l’imagerie de contraste de phase des rayons X
Résumé
In-line phase contrast imaging can be obtained with X-rays if the beam is sufficiently coherent. Retrieving the phase and absorption
from one or several of these images is a non-linear, ill-posed inverse problem, however. Here, we propose a deep learning method to solve this
twofold problem from a single intensity measurement. It is based on the unrolling approach, in which an iterative algorithm is "unrolled" into a
sequence, and some part of the iterations is replaced with a convolutional neural network. An advantage is that no regularization is needed, only
the gradient of the data fidelity term is calculated at each iteration. The network is trained to learn a complete iteration from this information and
the current iteration. We apply the method to a gradient descent scheme using smooth total variation regularization (GD-TVϵ) and compare the
standard and the unrolling version (DUGD). We show that the unrolled version gives better results quantitatively and qualitatively, and that the
computation time is much shorter than the classical iterative version.
Origine | Fichiers produits par l'(les) auteur(s) |
---|