On the use of machine learning for solving computational imaging problems

Abstract

It has recently been recognized that compressed sensing, especially dictionaries and related methods, formally map to machine learning architectures, e.g. recurrent neural networks. This has led to rapid growth in algorithms and methods based on deep neural networks (but not only) for solving a variety of inverse and computational imaging problems. In this paper, we review these developments in the specific context of quantitative phase imaging and emphasizing the impact of object power spectral density and noise properties on the quality of the reconstructions.