Meta-Album MD_MIX: 800 Synthetic Character Images for Few-Shot OCR

OmniPrint-MD-mix is a synthetic dataset of 800 character images across 20 categories, created by Haozhe Sun in 2021. Images are 128x128 RGB and were generated using the OmniPrint synthesizer with specific nuisance parameters like perspective transformation and rotation. The dataset is released under a CC BY 4.0 license as part of the Meta-Album collection for few-shot learning.

Use Cases

• Benchmarking few-shot image classification models based on the 20-class structure.
• Training models for character recognition under synthetic distortions based on described parameters like rotation and shear.
• Researching domain generalization in OCR using a controlled, synthetic data source.
• Developing data augmentation strategies by studying the effects of pre-defined synthesis parameters.

Strengths

• Contains 800 images with precisely documented synthesis parameters, enabling reproducible experiments.
• Images are uniformly sized at 128x128 pixels, which likely simplifies preprocessing.
• Released under a permissive CC BY 4.0 license, facilitating open use and redistribution.

Limitations

• All images are synthetic, which may limit model performance on real-world, natural character images.
• Description metadata is limited; actual data quality requires manual inspection after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.

Provenance

Source

OmniPrint synthesizer, using photos taken by a personal mobile phone as source textures.

Collection Method

Synthetic generation via the OmniPrint configurable character synthesizer.