Generative DH-MBIR for real-time coherent imaging through turbulence

Abstract

Digital Holography Model-Based Iterative Reconstruction (DHMBIR) jointly estimates atmospheric phase and target reflectance from noisy holography frames via Expectation-Maximization (EM). High wind speeds decorrelate consecutive frames fast enough to trap deterministic EM in local minima, and real-time wavefront correction allows only a narrow compute window per frame. No existing method delivers both real-time speed and resistance to this decorrelation-driven failure. We propose Generative DHMBIR (GDHMBIR): it perturbs the reflectance estimate with unbiased, multiplicative exponential noise that preserves positivity and drives the state estimate out of local minima. On synthetic turbulence data, GDHMBIR reaches an average Strehl ratio of 0.90 using one EM iteration and two Maximization-step (M-Step) updates, sustaining a mean throughput above 10 kHz; at this operating point, phase and reflectance estimates converge to ground truth within a few hundred frames.