Towards practical single-shot phase retrieval with physics-driven deep neural network.

Area of Science:

• Optics
• Image Processing
• Machine Learning

Background:

• Phase retrieval (PR) is crucial for recovering complex signals from intensity measurements in optical imaging.
• Current deep learning methods for single-shot PR struggle with accuracy due to domain disparity.
• Traditional physics-informed iterative methods are slow and may not handle complex structures or real-world errors.

Purpose of the Study:

• To develop a novel, accurate, and efficient deep learning approach for single-shot phase retrieval.
• To improve reconstruction accuracy by incorporating physical constraints into a neural network architecture.
• To address limitations of existing methods, including speed and handling of practical system errors.

Main Methods:

• Proposed a physics-driven, multi-scale deep neural network (DNN) architecture named PPRNet.
• PPRNet utilizes a feedforward, end-to-end trainable structure.
• The network is guided by Fourier intensity measurements at multiple scales to enhance accuracy.

Main Results:

• PPRNet achieves higher reconstruction accuracy compared to traditional learning-based PR methods.
• The proposed method demonstrates superior speed due to its feedforward, non-iterative nature.
• Experimental validation on an optical platform confirms the practicality and effectiveness of PPRNet.

Conclusions:

• PPRNet offers a significant advancement in single-shot phase retrieval for optical imaging.
• The physics-driven, multi-scale DNN approach provides a faster and more accurate solution than existing methods.
• PPRNet shows strong potential for practical applications in optical systems.