Computational imaging through deep learning

Abstract

Computational imaging (CI) is a class of imaging systems that uses inverse algorithms to recover an unknown object from the physical measurement. Traditional inverse algorithms in CI obtain an estimate of the object by minimizing the Tikhonov functional, which requires explicit formulations of the forward operator of the physical system, as well as the prior knowledge about the class of objects being imaged. In recent years, machine learning architectures, and deep learning (DL) in particular, have attracted increasing attentions from CI researchers. Unlike traditional inverse algorithms in CI, DL approach learns both the forward operator and the objects' prior implicitly from training examples. Therefore, it is especially attractive when the forward imaging model is uncertain (e.g. imaging through random scattering media), or the prior about the class of objects is difficult to be expressed analytically (e.g. natural images).

In this thesis, the application of DL approaches in two different CI scenarios are investigated: imaging through a glass diffuser and quantitative phase retrieval (QPR), where an Imaging through Diffuser Network (IDiffNet) and a Phase Extraction Neural Network (PhENN) are experimentally demonstrated, respectively. This thesis also studies the influences of the two main factors that determine the performance of a trained neural network: network architecture (connectivity, network depth, etc) and training example quality (spatial frequency content in particular). Motivated by the analysis of the latter factor, two novel approaches, spectral pre-modulation approach and Learning Synthesis by DNN (LS-DNN) method, are successively proposed to improve the visual qualities of the network outputs. Finally, the LS-DNN enhanced PhENN is applied to a phase microscope to recover the phase of a red blood cell (RBC) sample.

Furthermore, through simulation of the learned weak object transfer function (WOTF) and experiment on a star-like phase target, we demonstrate that our network has indeed learned the correct physical model rather than doing something trivial as pattern matching.