Fourier ptychographic microscopy with untrained deep neural network priors

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning

Background:

• Fourier ptychographic microscopy (FPM) enables high-resolution imaging.
• Traditional FPM reconstruction often relies on iterative algorithms and large labeled datasets for deep learning methods.
• Aberrations and illumination fluctuations can degrade image quality in FPM.

Purpose of the Study:

• To develop a physics-assisted deep neural network scheme for high-resolution image reconstruction in FPM.
• To overcome the limitations of traditional training-based deep learning approaches in FPM.
• To improve image quality and robustness against aberrations and illumination variations.

Main Methods:

• Proposed a Fourier ptychographic microscopy using untrained deep neural network priors (FPMUP) scheme.
• Integrated parallel neural networks for sample function (amplitude and phase), pupil function, and illumination intensity.
• Optimized neural network parameters to fit experimentally measured low-resolution images without prior training data.

Main Results:

• Achieved high-resolution image reconstruction from multiple low-resolution images.
• Demonstrated superior image quality compared to traditional iterative algorithms, particularly in phase recovery.
• Successfully predicted the Fourier spectrum outside the synthetic aperture, eliminating ringing artifacts.

Conclusions:

• The FPMUP scheme offers a powerful, training-free approach for high-resolution FPM.
• The method enhances robustness against aberrations and illumination fluctuations.
• Future work could further improve resolution by accurately predicting sample Fourier spectra beyond the synthetic aperture.