Fully self-supervised physics-aware holographic depth estimation

Autofocusing is a well-studied topic in holography, with a wide range of proposed methods, from mathematical models to more recent learning-based approaches. However, there is no consensus on a universal method that can autofocus an input hologram independently of the experimental setup. Minor changes in factors such as wavelength, pixel pitch, or hologram resolution can drastically impact the autofocus outcome. In this paper, we introduce a universal methodology that adheres to the general framework of holographic autofocusing while eliminating the need for manual hyperparameter tuning and offering robust adaptability to diverse input data. Our approach autonomously extracts optimal numerical reconstruction distances, performs volumetric rendering of the hologram, and estimates the underlying scene geometry to achieve precise autofocusing. To ensure accuracy, the generated depth estimates are constrained by matching the ground-truth values through an iterative hologram regeneration process. Our method demonstrates superior robustness and generalization on both synthetic computer-generated holograms and optically acquired on-axis phase-shifting holograms, marking a significant step toward universal autofocusing in holography.