Deep plug-and-play HIO approach for phase retrieval.

Area of Science:

• Computational imaging
• Image reconstruction
• Applied optics

Background:

• Phase retrieval aims to reconstruct images from intensity measurements, a challenging nonlinear and ill-posed problem.
• Traditional methods often struggle with noise and initialization.
• Learning-based approaches offer promising alternatives for inverse problems.

Purpose of the Study:

• To present the mathematical framework for a novel plug-and-play phase retrieval method.
• To evaluate the performance of this learning-based approach against existing methods.
• To demonstrate the method's effectiveness in image quality, computational speed, and robustness.

Main Methods:

• Incorporation of learning-based priors into Gerchberg-Saxton type algorithms via plug-and-play regularization.
• Mathematical derivation of analytical update steps using half-quadratic splitting.
• Extensive performance evaluation through simulations on a large dataset.

Main Results:

• The plug-and-play method achieves state-of-the-art performance in phase retrieval.
• Demonstrated superior image quality and computational efficiency compared to other methods.
• Exhibited robustness to varying initializations and noise levels.

Conclusions:

• The proposed learning-based plug-and-play regularization is effective for phase retrieval.
• This method offers a powerful and efficient solution for recovering images from intensity-only data.
• The approach shows significant potential for various imaging applications.