Single-image super-resolution via regularized extreme learning regression for imagery from microgrid polarimeters

Date of Award

2017

Degree Name

M.S. in Electrical and Computer Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Advisor: Vijayan K. Asari

Abstract

Division of focal plane imaging polarimeters have the distinct advantage of being capable of obtaining temporally synchronized intensity measurements across a scene; however, they sacrifice spatial resolution in doing so due to their spatially modulated arrangement of the pixel-to-pixel polarizers and often result in aliased imagery. This shortcoming is often overcome through advanced demosaicing strategies that minimize the effects of false polarization while preserving as much high frequency content as possible. While these techniques can yield acceptable imagery, they tend to be computationally complex and the spatial resolution is often reduced below the native capabilities of the focal plane array. This thesis proposes a super-resolution method based upon a previously trained regularized extreme learning regression (RELR) that aims to recover missing high-frequency content beyond the spatial resolution of the sensor and correct low-frequency content, while maintaining good contrast between polarized and unpolarized artifacts presented in this thesis. For each of the four channels of the image, the modified RELR predicts the missing high-frequency and lowfrequency components that result from upsampling. These missing high-frequency components are then refined with a high pass filter and added back to the upsampled image. This provides a fast and computationally simple way of recovering missing high frequency components that are lost with current state-of-the-art demosaicing algorithms. The modified RELR provides better results than other visible band single-image super-resolution techniques and is much faster, thus making it applicable to real-time applications. The obtained results demonstrate the effectiveness of the modified RELR for a truth scenario (no aliasing resulting from undersampling) and a derived microgrid scenario (aliasing resulting from undersampling). The truth scenario shows that the modified RELR performs exceptionally better than other algorithms, however, the derived microgrid scenario demonstrates the problems that result from aliasing for single-image super-resolution algorithms. In general, for the degree of linear polarization (DoLP) image product, aliasing greatly distorts objects within a scene and none of the super-resolution algorithms could do anything to correct for it. The modified RELR showed superior performance against other super-resolution algorithms investigated at maintaining contrast between the polarized and unpolarized artifacts, which is of great importance. Future work is dedicated to coming up with fast ways to handle aliasing that is present in true microgrid imagery.

Keywords

Machine learning, Digital images Deconvolution, Image reconstruction, Polariscope, Electrical Engineering, Engineering, single image super resolution, microgrid polarimeter, machine learning, extreme learning machine

Rights Statement

Copyright © 2017, author

Link to Full Text

Share

COinS