PhENN is a convolutional deep neural network that reconstructs quantitative phase images from diffracted intensity measurements some distance away from the phase objects. PhENN is trained on known phase-intensity pairs created from a particular database (e.g. ImageNet) but then found to perform well on objects created from other databases (e.g. Faces-LFW, MNIST, etc.). In this paper, we analyze the dependence of quantitative phase measurement quality on PhENN's architecture and the layout of the lensless imaging system, in particular, the number of layers (depth), the size of the innermost layer (waist size), the presence or absence of skip connections, the choice of training loss function and the free space propagation distance.

Date Published: 4 March 2019
PDF: 8 pages
Proc. SPIE 10887, Quantitative Phase Imaging V, 108870T (4 March 2019); doi: 10.1117/12.2513310

Published in SPIE Proceedings Vol. 10887:
Quantitative Phase Imaging V
Gabriel Popescu; YongKeun Park, Editor(s)