Rahvusarhiiv/ArchXAI_textline_OCR

🔍 ArchXAI Estonian OCR Dataset

Dataset Description

This dataset contains 42,901 annotated text regions from Estonian historical documents, designed for optical character recognition (OCR) and document analysis tasks. The dataset includes images of text regions along with their corresponding transcriptions, and source documents.

📊 Dataset Summary

• Total Examples: 42,901
• Language: 🇪🇪 Estonian
• Dataset Size: ~8 GB
• Task: OCR, Text Detection, Document Analysis
• Domain: Historical Documents, Archival Materials

🗂️ Dataset Structure

📋 Features

• image: Image data containing text regions
• text: Ground truth transcription of the text in the image
• document: Source document name
• reference: AIS reference number

💻 Usage Examples

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Rahvusarhiiv/ArchXAI_textline_OCR")

# Access a sample
sample = dataset['train'][0]
image = sample['image']
text = sample['text']

# Filter by document
ordered_text = dataset['train'].sort('document')

🎯 Use Cases

This dataset is valuable for:

• OCR Model Training: Training optical character recognition systems for Estonian text
• Document Layout Analysis: Understanding document structure and reading order
• Historical Document Digitization: Processing Estonian archival materials
• Text Detection: Locating text regions in document images
• Multilingual OCR: Estonian language text recognition research
• Digital Humanities: Analysis of Estonian historical documents

⚠️ Data Quality Notes

• Images contain text regions extracted from historical Estonian documents
• Text transcriptions follow modern Estonian orthography where applicable
• Some historical texts may contain archaic spellings or formatting
• Coordinate systems may vary depending on source document resolution

🚫 Limitations

• Limited to Estonian language content
• Historical document quality may affect image clarity
• Transcription accuracy depends on document preservation state
• Dataset size is relatively small (42,901 examples)

📞 Contact

Depending on the nature of your question, please contact one of the following:

• Content of dataset: @svlp or @LudwigRoine
• Format or anything technical: @paulpall
• For everything else: the National Archives of Estonia.