Real-time phase imaging with physics-enhanced network and equivariance.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning for Scientific Applications

Background:

• Supervised learning for phase imaging demands extensive, often unattainable, datasets.
• Existing methods struggle to balance high fidelity with rapid image reconstruction.

Purpose of the Study:

• To introduce a novel unsupervised learning framework for real-time phase imaging.
• To leverage physical principles for efficient and accurate phase reconstruction from single diffraction patterns.

Main Methods:

• Developed the Physics-Enhanced Network and Equivariance (PEPI) architecture.
• Incorporated measurement and equivariant consistency for network optimization.
• Utilized a total variation kernel (TV-K) regularization for enhanced detail.

Main Results:

• PEPI reconstructs object phase rapidly and accurately from single diffraction patterns.
• The unsupervised PEPI method demonstrates performance comparable to supervised techniques.
• PEPI excels at preserving high-frequency details and texture information.

Conclusions:

• PEPI offers a robust and generalizable solution for unsupervised phase imaging.
• The physics-informed approach significantly improves performance on imaging inverse problems.
• This work enables high-precision unsupervised phase imaging, overcoming data limitations.