Fourier phase retrieval using physics-enhanced deep learning.

Area of Science:

• Computational imaging
• Applied physics
• Deep learning

Background:

• Fourier phase retrieval (FPR) reconstructs images from Fourier transform magnitudes, crucial in many scientific and engineering fields.
• The ill-posed nature of FPR presents significant challenges for accurate image reconstruction.
• Existing methods often struggle with stability and accuracy due to the inherent complexity of the problem.

Purpose of the Study:

• To propose a novel learning-based approach for Fourier phase retrieval (FPR).
• To integrate the physical model of the FPR imaging system with deep neural networks.
• To enhance the accuracy and stability of image reconstruction in FPR applications.

Main Methods:

• A two-step learning-based method combining self-supervised data generation and physics-informed fine-tuning.
• Leveraging the image formation model for self-supervised training data generation.
• Exploiting the physical model to enforce physics-consistency constraints on network predictions.

Main Results:

• The proposed method successfully integrates implicit priors from training data and explicit priors from the physical imaging model.
• Simulations and experimental results demonstrate high accuracy and stability in Fourier phase retrieval.
• The approach effectively addresses the ill-posed nature of the FPR problem.

Conclusions:

• The developed physics-informed deep learning approach offers a robust solution for Fourier phase retrieval.
• The method shows significant potential for wide application across various scientific and engineering disciplines utilizing FPR.
• Source code is available for non-commercial use, promoting further research and development.