Simulation-Trained Sparse Coding for High-Precision Phase Imaging in Low-Dose Electron Holography.

Area of Science:

• Materials Science
• Machine Learning
• Electron Microscopy

Background:

• Low-dose electron holography is crucial for imaging beam-sensitive materials.
• Shot noise in low-dose electron holography significantly degrades image quality and phase information.
• Existing denoising methods may struggle with extremely low signal-to-noise ratios.

Purpose of the Study:

• To adapt and validate sparse coding for noise reduction in low-dose electron holography.
• To develop a dictionary of interference fringe features from simulated holograms.
• To assess the performance of simulation-trained sparse coding on both simulated and experimental holograms.

Main Methods:

• Generated simulated electron holograms with realistic shot noise.
• Trained a sparse coding dictionary using these simulated holograms.
• Applied the learned dictionary to denoise simulated and experimental holograms with varying signal-to-noise ratios (SNRs).

Main Results:

• Successfully removed noise from holograms with SNRs as low as 0.10.
• Denoised holograms accurately represented phase distributions.
• The simulation-trained dictionary effectively denoised experimental holograms from a p-n junction specimen.

Conclusions:

• Simulation-trained sparse coding is a robust method for denoising low-dose electron holograms.
• This approach is applicable across a wide range of imaging conditions, including for electron beam-sensitive materials.
• The method offers accurate phase retrieval even at extremely low SNRs.