Phase retrieval with physics informed zero-shot network.

Area of Science:

• Computational imaging
• Deep learning applications
• Phase retrieval algorithms

Background:

• Accurate phase estimation from diffracted intensity images relies on prior information.
• Deep learning offers high-quality reconstruction but demands extensive labeled datasets for training.
• Existing methods face challenges due to data requirements for deep learning models.

Purpose of the Study:

• To propose a zero-shot learning method for phase retrieval.
• To demonstrate the utility of a denoising-trained neural network for phase retrieval.
• To reduce the data dependency in deep learning-based phase retrieval.

Main Methods:

• A zero-shot learning framework is proposed for phase retrieval.
• A deep neural network, trained for denoising, is adapted for phase retrieval by enforcing diffraction physics.
• Total variation is incorporated to enhance reconstruction efficiency.

Main Results:

• The proposed method successfully retrieves phase information without large labeled datasets.
• The denoising-trained network effectively serves as an object prior for phase retrieval.
• Incorporating total variation accelerated reconstruction by approximately ninefold with comparable quality.

Conclusions:

• Zero-shot learning offers a viable solution to data requirements in deep learning for phase retrieval.
• Denoising priors can be repurposed for phase retrieval tasks, reducing training data needs.
• The integration of total variation significantly improves the speed of phase retrieval.