4D wavefront sensing using dynamic neural image sharpening.

Abstract

In this paper, we develop a dynamic framework for wavefront sensing and image sharpening from a video sequence of data collected by a digital-holographic sensor. We build on recent progress using implicit neural representations (INRs) to model the complex phase errors caused by atmospheric turbulence. INRs leverage unsupervised learning to form a continuous functional representation of the atmospheric phase errors. In turn, we extend our framework from modeling 3D phase errors from a single digital-holographic measurement (consisting of three spatial coordinates, x, y, and z), to modeling 4D phase errors from a time sequence of digital-holographic measurements (consisting of three spatial coordinates plus time, x, y, z, and t). By introducing additional time-varying measurements, we better model the dynamic behavior of atmospheric turbulence. This improved modeling provides three benefits: (1) it helps constrain our under-determined estimation problem, which improves phase-error estimation quality and image reconstruction; (2) it allows us to estimate phase errors between actual measurements, increasing the effective sampling rate of our sensor; and (3) it allows us to predict future phase errors, helping us reduce latency in our estimation pipeline. At large, (1)-(3) creates numerous benefits for applications where 4D phase errors naturally arise (e.g., microscopy, metrology, lidar, remote sensing, etc.).