Trent Kyono, The Boeing Company; Jacob Lucas, The Boeing Company; Michael Werth, The Boeing Company; Justin Fletcher, SMC/DirSP-G; Ian McQuaid, Air Force Maui Optics and Supercomputing Site
Keywords: Machine Learning, Deep neural networks, convolutional neural networks
Abstract:
Imaging space objects from ground-based stations results in images perturbed with atmospheric turbulence. To reconstruct the original image from an observed perturbed image, Multi-frame Blind Deconvolution (MFBD) is a state-of-the-art maximum-likelihood algorithm that in practice is often computationally expensive in terms of time and space. It may take on the order on hours to generate reconstructions and typically requires hours of compute time on a super computer. Even GPU based approaches currently take over an hour to converge which is not fast enough for real-time situational awareness. We present preliminary results on neural network reconstructions for solving this inverse imaging problem. We investigate several architectures and techniques that are trained on a large synthetic dataset and applied to real data collected at Maui Space Surveillance Site.
Date of Conference: September 15-18, 2020
Track: Adaptive Optics & Imaging