cGAN-assisted imaging through stationary scattering media

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Image analysis through scattering media is difficult due to speckle decorrelations.
• In-line imaging modalities are compact and robust but face challenges negating scattering effects.
• Digital holographic microscopy is an in-line technique where scattering is a significant issue.

Purpose of the Study:

• To explore the use of conditional generative adversarial networks (cGANs) for mitigating scattering in in-line imaging.
• To assess the effectiveness of cGANs in handling additional scatterers in digital holographic microscopy.
• To determine if cGANs can enable quantitative feature extraction from scattered images.

Main Methods:

• Utilized light scattering simulations and experimental setups.
• Employed conditional generative adversarial networks (cGANs) for image restoration.
• Trained cGANs with minuscule datasets (as few as 20 image pairs).

Main Results:

• cGANs were rapidly trained and efficiently learned the statistical mapping between scattered and clear images.
• Output images were sufficiently faithful for quantitative feature extraction.
• Accurate localization of diffraction-limited impulses was achieved, transforming shift-variant to linear shift-invariant systems.

Conclusions:

• cGANs offer a powerful solution for scattering mitigation in in-line imaging.
• The method is efficient, requiring minimal data and training time.
• This approach enhances the capabilities of compact imaging systems for precise analysis.