Michael Werth, The Boeing Company; Brandoch Calef, The Boeing Company; Kevin Roe, The Boeing Company; Amanda Conti, Air Force Research Laboratory
Keywords: Imaging, Ground-Based Telescopes, GPU, GPGPU, CUDA, Software
Abstract:
Ground-based observations of Low Earth Orbit (LEO) satellites are significantly degraded by atmospheric turbulence. Multi-Frame Blind Deconvolution (MFBD) is a class of algorithm that attempts to solve this ill-posed inverse problem using a forward model approach. It assumes that a Point Spread Function (PSF) is convolved with a pristine image of the target and then iteratively solves for both simultaneously. Processing one full LEO collection in this way requires an extraordinary amount of processing power and is a task that is normally performed with supercomputers to minimize processing time. In this paper, we describe a new MFBD implementation written with NVidias Compute Unified Device Architecture (CUDA) language to make use of the incredible parallelization capabilities of graphical processing units (GPUs), advancing the state-of-the-art for image resolvability. Using GPUs reduces the scope of hardware achieving the same processing time from a supercomputer-scale system to a single desktop or server. Measurements of processing time and image quality for several reference objects are compared against results from the latest version of Physically Constrained Iterative Deconvolution (PCID), a gold-standard MFBD implementation that uses the Message Passing Interface (MPI) to make use of numerous CPUs in high-performance computing environments.
Date of Conference: September 17-20, 2019
Track: Adaptive Optics & Imaging