Randomized probe imaging through deep k-learning.

Area of Science:

• Diffractive Imaging
• Computational Imaging
• Machine Learning Applications

Background:

• Randomized probe imaging (RPI) is a single-frame technique for reconstructing scattering objects.
• Phase retrieval in RPI typically relies on slow iterative algorithms.
• Current methods struggle with large datasets and photon-starved conditions.

Purpose of the Study:

• To develop a fast and efficient deep learning method for phase retrieval from RPI data.
• To improve the computational efficiency and robustness of RPI reconstructions.
• To enable analysis of larger datasets and support studies of dynamic phenomena.

Main Methods:

• Proposed 'deep k-learning', a novel deep learning approach for RPI.
• Utilized the physical propagation operator to generate initial object approximations for the neural network.
• Trained the network to bypass parametrization of far-field diffraction physics.

Main Results:

• Deep k-learning demonstrated significant computational efficiency.
• The method proved robust against Poisson noise.
• Reconstruction quality was dramatically improved compared to traditional iterative methods.

Conclusions:

• Deep k-learning offers a computationally efficient and robust solution for RPI phase retrieval.
• The method facilitates analysis of large datasets, particularly in photon-starved scenarios.
• Potential applications include studying dynamic phenomena in physical sciences and biological engineering.