Meta-Album DIBaS Mini: 33 Bacterial Species Images for Few-Shot Learning

A preprocessed version of the Digital Images of Bacteria Species (DIBaS) dataset, containing 1,320 images across 33 bacterial species. The images were split and resized to 128x128 pixels and processed with blob normalization to reduce color bias. The dataset was created for the Meta-Album benchmark by Romain Mussard in March 2022.

Use Cases

• Benchmarking few-shot learning algorithms on microscopic imagery based on the 33 bacterial species classes.
• Training image classifiers for bacterial species identification based on preprocessed 128x128 images.
• Studying the effect of blob normalization on model bias and decision-making in image classification tasks.

Strengths

• Contains 1,320 images with at least 108 images per class, providing a balanced foundation for classification tasks.
• Images are preprocessed to a consistent 128x128 pixel size, which may simplify model input pipelines.
• Applies blob normalization, a technique described as unusual for this image type, to address color bias.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• The original large images were split, which may affect the biological context or features present in each sample.
• Row count is unknown, which may limit suitability assessment for specific model scales.

Provenance

Source

Digital Image of Bacterial Species (DIBaS) dataset, preprocessed for the Meta-Album benchmark.

Collection Method

Original microscopic images were split into smaller images and resized.