SEM Image Collections with AI Models and Quality Metrics

Six collections of Scanning Electron Microscope images are provided, augmented with Poisson noise and contrast variations. The dataset includes three trained U-Net AI models for segmentation, alongside CSV files containing image quality metrics and model accuracy evaluations. This work was funded by the CHIPS Metrology Program and published by the National Institute of Standards and Technology in July 2025.

Use Cases

• Benchmarking segmentation model accuracy using the provided AI model accuracy metrics CSV files.
• Analyzing the impact of Poisson noise and contrast augmentation on segmentation performance across the six image collections.
• Training new segmentation models using the provided U-Net models as a starting point on the SEM images.
• Evaluating image quality metrics against ground truth mask images to establish detection limits.

Strengths

• Includes three pre-trained TensorFlow AI models for immediate application.
• Provides six distinct collections of SEM images with controlled noise and contrast augmentations.
• Contains separate CSV files for structured analysis of image quality and model accuracy metrics.

Limitations

• The total number of images and rows in the CSV files is unknown, hindering size estimation.
• Specific performance details like class balance or model precision/recall are not detailed in the description.
• The dataset's focus on augmented SEM images may limit direct application to raw, unprocessed microscopy data.

Provenance

Source

National Institute of Standards and Technology

Collection Method

Created with funding from the CHIPS Metrology Program; includes generated images, trained models, and calculated metrics.

Freshness

Last updated July 2025.