LUCID: Accelerating Image Reconstructions of LEO Satellites Using GPUs

Abstract

We report on a new Multi-Frame Blind Deconvolution (MFBD) implementation developed to reconstruct high resolution images of Low Earth Orbit (LEO) satellites from short-exposure ensembles of images recorded from a large-diameter ground-based telescope. This implementation, which is named Likelihood-based Uncertainty-Constrained Iterative Deconvolution (LUCID), uses NVidia's Compute-Unified Device Architecture (CUDA) to perform iterative blind deconvolution processing on Graphical Processing Units (GPUs). A single instance of LUCID is capable of using multiple GPUs to segment and process a large collection of image frames in parallel, achieving significant reductions in processing time and hardware cost compared to equivalent algorithms that only use Central Processing Units (CPUs). In this paper we describe how LUCID makes use of CUDA and GPU hardware to produce image reconstructions. We also show performance comparisons between the gold-standard Physically Constrained Iterative Deconvolution (PCID) implementation of MFBD running on all of the cores of an enterprise-class CPU and LUCID running on various target GPU platforms, including consumer-grade and datacenter-grade devices. The quality of image reconstructions produced by LUCID is demonstrated by showing data collected with the 3.6-meter Advanced Electro-Optical System (AEOS) telescope before and after processing.