Multi-dimensional computational imaging from diffraction intensity using deep neural networks

Abstract

Diffraction of light can be found everywhere in nature, from sunlight rays fanning out from clouds to multiple colors reflected from the surface of a CD. This phenomenon of light explains any change in the path of light due to an obstacle and is of particular significance as it allows us to see transparent (or pure-phase) objects, e.g. biological cells under visible-wavelength light or integrated circuits under X-rays, with proper exploitation of the phenomenon. However, cameras only measure the intensity of the diffracted light, which makes the camera measurements incomplete due to the loss of phase information. Thus, this thesis addresses the reconstruction of multi-dimensional phase information from diffraction intensities with a regularized inversion using deep neural networks for two- and three-dimensional applications. The inversion process begins with the definition of a forward physical model that relates a diffraction intensity to a phase object and then involves a physics-informing step (or equivalently, physics prior) to deep neural networks, if applicable. In this thesis, two-dimensional wavefront aberrations are retrieved for high-contrast imaging of exoplanets using a deep residual neural network, and transparent planar objects behind dynamic scattering media are revealed by a recurrent neural network, both in an end-to-end training fashion. Next, a multi-layered, three-dimensional glass phantom of integrated circuits is reconstructed under the limited-angle phase computed tomography geometry with visible-wavelength laser illumination using a dynamical machine learning framework. Furthermore, a deep neural network regularization is deployed for the reconstruction of real integrated circuits from far-field diffraction intensities under the ptychographic X-ray computed tomography geometry with partially coherent synchrotron X-ray illumination.