Enhancing Semantic Segmentation in High-Resolution TEM Images: A Comparative Study of Batch Normalization and...

Area of Science:

• Materials Science
• Nanotechnology
• Computer Vision

Background:

• Deep learning (DL) enhances transmission electron microscopy (TEM) image analysis for materials science.
• DL automates feature detection and accelerates data analysis in complex TEM datasets.
• Accurate nano- and microscale characterization is vital for nanoparticle research.

Purpose of the Study:

• Investigate the impact of batch normalization (BN) versus instance normalization (IN) on DL model performance for TEM image semantic segmentation.
• Compare BN and IN in U-Net and ResNet architectures for high-resolution TEM image analysis.
• Determine the optimal normalization technique for DL-based TEM image analysis.

Main Methods:

• Trained U-Net and ResNet models using both BN and IN.
• Evaluated model performance on two distinct high-resolution TEM image datasets.
• Utilized semantic segmentation for precise feature identification.

Main Results:

• Instance normalization (IN) consistently outperformed batch normalization (BN).
• IN achieved higher Dice scores and Intersection over Union (IoU) metrics.
• The choice of normalization significantly impacts DL model performance on TEM images.

Conclusions:

• Instance normalization is superior to batch normalization for deep learning-based TEM image semantic segmentation.
• Selecting the appropriate normalization method is crucial for maximizing DL model performance in materials science applications.
• This research provides essential insights for optimizing DL workflows in electron microscopy.