Physics‐Consistent Dual‐Plane‐Driven Complex‐Domain Learning for Phase Retrieval in Inline Holography

Abstract

Inline holography enables non‐invasive and label‐free quantitative phase imaging (QPI) with a compact and stable optical configuration. However, it remains hindered by severe twin‐image artifacts due to the absence of sensor‐plane phase information. Iterative approaches are sensitive to noise and prone to local minima, while existing learning‐based approaches rarely exploit sensor‐plane phase as a physically meaningful initialization and insufficiently integrate diffraction physics with complex‐domain learning. Here, we propose a dual‐plane‐driven complex‐valued cascaded neural network (DP‐CVNN) for phase retrieval in inline holography. The framework combines a front‐end real‐valued network for sensor‐plane phase estimation with a back‐end complex‐valued neural network for diffraction propagation learning, thereby enabling physically meaningful reconstruction from single‐shot inline holograms. By jointly constraining sensor‐plane and object‐plane phases through diffraction‐consistent supervision, the proposed framework effectively bridges data‐driven holographic reconstruction with physics‐consistent optical modeling in the complex domain, thereby reducing sensitivity to noise‐contaminated intensity measurements. Simulations and experiments on diverse phase‐dominant samples demonstrate effective suppression of twin‐image artifacts, achieving favorable reconstruction fidelity and noise robustness compared with representative conventional and learning‐based methods. Our findings further validate the critical role of sensor‐plane phase initialization, as well as its potential to benefit conventional phase retrieval algorithms.