Deep learning-based noise filtering toward millisecond order imaging by using scanning transmission electron...

Area of Science:

• Materials Science
• Electron Microscopy
• Data-Driven Science

Background:

• Scanning Transmission Electron Microscopy (STEM) is crucial for data-driven materials science.
• Current STEM temporal resolution limits real-time observation of dynamic phenomena.
• Rapid image acquisition in STEM suffers from distortion, noise, and blurring.

Purpose of the Study:

• To develop an image correction framework for rapid STEM imaging.
• To improve STEM's temporal resolution for in situ observations.
• To enable high-quality, video-rate STEM without hardware modifications.

Main Methods:

• Integration of deep learning (DL) for denoising and distortion correction.
• Optimization of DL schemes for rapid STEM image acquisition.
• Comparison of corrected rapid scan images with regular speed scan images.

Main Results:

• The DL network effectively removes statistical noise and unidirectional blurring.
• Rapid STEM imaging achieved high quality comparable to regular speed imaging.
• Demonstrated feasibility of video-rate STEM observation through software-based correction.

Conclusions:

• Deep learning provides a viable solution for high-quality, rapid STEM imaging.
• The DL-based approach enables enhanced in situ observations of dynamic material processes.
• This method advances STEM capabilities for real-time materials characterization.