PhaseGAN: a deep-learning phase-retrieval approach for unpaired datasets.

Area of Science:

• Computational imaging
• Optics and photonics
• Machine learning applications

Background:

• Deep learning (DL) phase retrieval offers real-time solutions but typically requires paired datasets and often neglects imaging physics.
• Existing DL methods are limited to scenarios with pre-existing phase solutions, hindering their application in novel or challenging experimental conditions.

Purpose of the Study:

• To introduce PhaseGAN, a new deep learning (DL) approach for phase retrieval using Generative Adversarial Networks (GANs).
• To overcome limitations of current DL methods by enabling the use of unpaired datasets and incorporating image formation physics.

Main Methods:

• Developed PhaseGAN, a Generative Adversarial Network (GAN) architecture for phase retrieval.
• Integrated image formation physics into the DL model.
• Introduced a novel Fourier loss function to enhance reconstruction accuracy.

Main Results:

• PhaseGAN successfully reconstructs phase information from intensity holograms or diffraction patterns.
• The method utilizes unpaired datasets, a significant advancement over existing techniques.
• PhaseGAN demonstrates robust performance, particularly in scenarios where conventional algorithms fail, such as ultra-fast experiments.

Conclusions:

• PhaseGAN provides a powerful, real-time phase retrieval solution by leveraging unpaired data and incorporating physical imaging principles.
• This approach expands the applicability of DL for phase retrieval, especially in complex and time-critical experimental settings.
• PhaseGAN offers a viable alternative when direct phase reconstructions are unavailable but simulations or related data exist.