Digital twin-inspired deep learning for high-fidelity reconstruction of digital holograms

Digital holography, which enables quantitative phase imaging, suffers from reconstruction degradation caused by environmental noise and system imperfections. We propose a digital twin-inspired deep learning framework for high-fidelity reconstruction of digital holograms, in which a physics-based simulation of the recording system is used to generate training data in a virtual domain. This approach alleviates the dependence on ideal optical setups and strictly controlled experimental conditions. We further investigate supervised, unsupervised, and self-supervised learning schemes for high-fidelity reconstruction of digital holograms recorded from phase objects. Experimental results demonstrate that the proposed framework improves reconstruction fidelity under realistic experimental conditions.