Variational Model-Based Deep Neural Networks for Image Reconstruction

In recent years, we have witnessed unprecedented growth of research interests in deep learning approaches to image reconstruction. A majority of these approaches are inspired by the well-developed variational method and associated optimization algorithms for the inverse problem of image reconstruction. These approaches mimic the iterative schemes of the standard optimization algorithms but integrate learnable components to form structured deep neural networks, and employ large amount of observation data to train the networks for the specific reconstruction tasks. They have demonstrated significantly improved empirical performance and requiremuch lower computational cost compared to the classical methods in a variety of applications. We provide the details of the derivations, the network architectures and the training procedures for several typical networks in this field.