Title

Deep Learning Approach to Structure from Polarization

Date of Award

2021

Degree Name

M.S. in Electrical and Computer Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Bradley Michael Ratliff

Abstract

Imaging polarimeters are a type of imaging device that attempts to estimate the polarized Stokes vector at each point in an imaged scene. Polarization has shown the ability to reduce background clutter, defeat atmospheric scatterers, improve scene contrast within polarized regions, and provide shape information about polarized objects of interest. Measured angle of polarization imagery tends to be highly correlated with the azimuthal component of object surface normal vectors. Hence, while polarimetric images do not directly provide 3D scene information, our goal in this work is to investigate the applicability of deep learning approaches for estimation of 3D structure from polarimetric image data. Unlike other image modalities, no repositories of polarimetric training data are readily available for training and testing purposes. As part of this work, we design a set of laboratory-based data collection experiments under a controlled set of scene conditions to obtain a sufficient set of polarimetric training and testing data. We then develop a deep learning approach to structure from polarization based upon the Pix2Pix conditional generative adversarial network for image translation problems. Initial results from training and testing our approach are presented that demonstrate promise for obtaining pixel-wise 3D information from polarimetric image data.

Keywords

Electrical Engineering, Optics, Computer Engineering, Engineering, Deep learning, polarization, 3D structure, Degree of linear polarization, Angle of polarization

Rights Statement

Copyright 2021, author

Link to Full Text
COinS