NIST Handprinted Character Images from 3600 Writers

NIST Special Database 19 contains a corpus of training materials for handprinted document and character recognition. It includes handprinted sample forms from 3600 writers and 810,000 isolated character images with ground truth classifications. The dataset also provides reference forms and software utilities for image management.

Use Cases

• Train a character recognition model on 810,000 isolated character images using the provided ground truth classifications.
• Analyze handwriting style variations across the 3600 writers represented in the handprinted sample forms.
• Benchmark OCR algorithms using the ground truth classifications associated with each character image.
• Develop image preprocessing pipelines using the provided software utilities for managing the character images.

Strengths

• Large-scale collection of 810,000 isolated handprinted character images.
• Data sourced from a diverse pool of 3600 individual writers.
• Includes ground truth classifications for supervised learning tasks.
• Provides reference forms and software utilities for data handling.

Limitations

• Image data is from handprinted forms, not natural handwriting, which may limit realism for some applications.
• The dataset consists of two different editions with images in different formats (MIS/PCT and PNG), requiring format awareness.
• No information is provided on the demographic or geographic distribution of the 3600 writers.

Provenance

Source

National Institute of Standards and Technology (NIST).

Collection Method

Collected from handprinted sample forms provided by 3600 writers.

Freshness

Last updated in June 2023.