A08 Sparse exit wave reconstruction via deep unfolding

The aim of this project is to introduce hybrid models for exit wave reconstruction, a problem from electron microscopy and a variant of phase retrieval, using deep unfolding and to understand the behavior of the resulting neural networks, e.g., their generalization properties. We systematically explore which network parameters are most beneficial for learning, considering different variants of data term and regularizer, focussing on sparsity inducing regularizers like the 1-norm. These are combined with a sparsity basis that is introduced as network parameter shared among the layers.