Leveraging unlabeled SEM datasets with self-supervised learning for enhanced particle segmentation.

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Scanning Electron Microscopes (SEMs) generate vast image data requiring extensive user analysis.
• Automating SEM image analysis is crucial for efficiency in experimental science.
• Machine Learning (ML), particularly supervised learning, is effective but hindered by manual annotation needs.

Purpose of the Study:

• To evaluate Self-Supervised Learning (SSL) techniques for automated SEM image analysis.
• To introduce and assess novel SSL methods using the ConvNeXtV2 architecture for particle detection.
• To provide insights into dataset size impact on SSL performance for practical applications.

Main Methods:

• Development of a framework to evaluate SSL techniques on SEM image data.
• Leveraging the ConvNeXtV2 architecture for particle detection in SEM images.
• Curating a dataset of 25,000 SEM images for benchmarking SSL methods.
• Performing an ablation study on dataset size and SSL performance.

Main Results:

• ConvNeXtV2-based SSL models demonstrated superior particle detection performance across various scales.
• Achieved up to a 34% reduction in relative error compared to existing SSL methods.
• Identified key relationships between dataset size and SSL model performance.

Conclusions:

• SSL, particularly with ConvNeXtV2, offers a powerful alternative to supervised learning for SEM image analysis.
• The proposed framework and findings facilitate the integration of SSL into autonomous analysis pipelines.
• This research accelerates materials science discovery by enhancing automated image analysis capabilities.