DH-GAN: a physics-driven untrained generative adversarial network for holographic imaging.

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Digital holography (DH) is a 3D imaging technique that reconstructs object shapes by analyzing recorded holograms.
• Supervised deep learning (DL) methods for DH require extensive datasets, which are often unavailable due to sample scarcity or privacy concerns.
• Existing one-shot DL methods for DH may neglect underlying physics, leading to non-explainable and non-generalizable results.

Purpose of the Study:

• To develop a novel deep learning architecture for accurate and explainable 3D reconstruction in digital holography.
• To overcome the limitations of data scarcity and the black-box nature of current DL-based DH methods.
• To improve the reconstruction quality and robustness to noise in digital holography.

Main Methods:

• Proposed a generative adversarial network (GAN) architecture for hologram formation inverse modeling.
• Integrated a discriminative network for semantic reconstruction quality assessment.
• Employed a progressive masking module with simulated annealing to enforce background smoothness and enhance reconstruction.

Main Results:

• Achieved significant improvements in reconstruction quality, with approximately 5 dB PSNR gain over competitor methods.
• Demonstrated enhanced robustness to noise, showing a 50% reduction in PSNR degradation rate.
• Exhibited high transferability to similar samples, enabling rapid deployment in time-sensitive applications without retraining.

Conclusions:

• The proposed physics-informed DL method offers a superior alternative for 3D reconstruction in digital holography, especially when training data is limited.
• The GAN-based approach provides explainable, generalizable, and transferable solutions for holographic imaging.
• This advancement facilitates practical applications of DH in various fields requiring efficient and accurate 3D imaging.