Noise2Atom: unsupervised denoising for scanning transmission electron microscopy images

Area of Science:

• Materials Science
• Image Processing
• Deep Learning

Background:

• Scanning Transmission Electron Microscopy (STEM) generates high-resolution images crucial for materials science.
• Image noise in STEM datasets degrades atomic resolution and hinders accurate analysis.
• Existing denoising methods often require paired noisy and clean images or specific noise models.

Purpose of the Study:

• To introduce Noise2Atom, an effective deep learning model for denoising STEM image series.
• To enable the mapping of noisy experimental STEM images to clear, atomic-resolution images.
• To develop a denoising solution that does not rely on signal priors, noise estimation, or paired training data.

Main Methods:

• Noise2Atom employs a deep learning architecture to map noisy STEM images to a clean domain.
• The model incorporates two external networks to integrate domain knowledge as constraints.
• A novel metric, consecutive structural similarity (CSS), is proposed for evaluating image restoration quality due to the absence of ground truth.

Main Results:

• Noise2Atom successfully denoises STEM image series, producing clear atomic images.
• The model demonstrates superior performance compared to existing methods, validated by CSS and visual quality assessments.
• Evaluations were conducted on diverse experimental STEM datasets, showcasing the model's generalizability.

Conclusions:

• Noise2Atom offers an effective, unsupervised deep learning approach for STEM image denoising.
• The model alleviates the need for paired data and complex noise characterization in STEM imaging.
• The proposed CSS metric provides a reliable method for assessing image restoration in the absence of ground truth.