Rapid eigenpatch utility classifier for image denoising.

Area of Science:

• Electron microscopy
• Image processing
• Materials science

Background:

• Low-illumination electron microscopy images suffer from Poisson and Gaussian noise.
• High dose rates improve signal-to-noise but damage delicate specimens.
• Existing denoising methods can reduce image contrast or introduce artifacts.

Purpose of the Study:

• To develop a lightweight denoising architecture that preserves experimental data integrity.
• To combine the strengths of patch-based and deep neural network (DNN) approaches for effective noise removal.
• To mitigate DNN-induced artifacts while maintaining denoising performance.

Main Methods:

• A hybrid approach combining non-local patch-based Singular Value Decomposition (SVD) with a convolutional neural network (CNN).
• The Rapid Eigenpatch Utility Classifier for Image Denoising (REUCID) uses SVD for component identification and CNN for classification.
• The CNN acts solely as a classifier on SVD eigenvectors, preventing overreach and artifact introduction.

Main Results:

• The REUCID method effectively removes noise from electron microscopy images.
• The hybrid approach preserves essential image contrast and genuine structural features.
• Demonstrated superior performance on high-angle annular dark-field images compared to conventional techniques.

Conclusions:

• The REUCID offers an effective and data-integrity-preserving solution for electron microscopy image denoising.
• This classification-only DNN integration represents a significant advance in mitigating artifacts.
• The method enhances image contrast and preserves structural details, crucial for material characterization.