Blind single-shot phase retrieval based on a self-supervised physics-adaptive neural network.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Single-shot phase retrieval reconstructs samples from single diffraction patterns.
• Existing methods require precise physical models and struggle with unknown diffraction distances.
• Blind single-shot phase retrieval is challenging due to unknown physical parameters.

Purpose of the Study:

• To develop a robust method for blind single-shot phase retrieval overcoming unknown diffraction distances.
• To introduce a self-supervised physics-adaptive neural network for accurate sample reconstruction.
• To enable self-calibrated snapshot coherent diffraction imaging.

Main Methods:

• A self-supervised physics-adaptive neural network (BlindPR-SSPANN) is proposed.
• The network jointly optimizes physical parameters and reconstruction network parameters.
• A novel architecture integrates tunable physical parameters in a multi-stage, coupled reconstruction process.
• Training utilizes a self-supervised scheme with a refined physics-consistent loss function.

Main Results:

• BlindPR-SSPANN achieves high-performance reconstructions from single intensity measurements.
• The method demonstrates robustness against significant diffraction distance errors.
• Successful self-calibrated snapshot coherent diffraction imaging was enabled.

Conclusions:

• BlindPR-SSPANN effectively addresses the challenges of blind single-shot phase retrieval.
• The proposed physics-adaptive approach enhances reconstruction accuracy and self-calibration.
• This work advances coherent diffraction imaging with improved robustness and applicability.