Andrew Vanden Berg, Air Force Research Lab; Ian Cunnyngham, MorphOptic, Inc.; Justin Fletcher, United States Space Force Space Systems Command
Keywords: Closely-spaced objects, machine learning, speckle interferometry
Abstract:
Speckle interferometry is a well-established astronomical technique that enables detection of closely-spaced objects (CSOs) near the theoretical limit of a telescope’s aperture. By utilizing this technique, the resolution-limiting effects of the atmosphere can be largely mitigated. Observations comprise many, extremely short exposures which effectively freeze the turbulence of the atmosphere. These frames are then stacked in the spatial-Fourier domain where the phase errors induced by the atmosphere cancel out with enough examples. If a CSO is present, a characteristic fingerprint can be observed in the amplitude of the stacked images. The structure of this fingerprint is determined by parameters of the underlying scene, including apparent radial separation and brightness of the objects. In certain limits of these parameters, this fingerprint degrades. There remains an unsolved decision problem, which is to determine the plurality of objects, given a speckle interferometric image. In this work, we formalize this decision problem as a binary classification task: SpeckleNet. We apply modern computer vision approaches, including convolutional neural networks, to classify images by plurality and characterize the performance of each classifier over a range of underlying parameters. The resulting models require only a few milliseconds to infer plurality and can be adapted to virtually any optical telescope, requiring only collection software and post-processing changes.
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications