corste1/MTJ-OCR

Early 20th Century Historical Document OCR Dataset

This is a large-scale, manually annotated image dataset specifically designed for early historical documents from the last century. It contains relevant historical literature from that period, complete with Optical Character Recognition (OCR) annotations.

Notably, the dataset is divided into two main parts:

• Simulated Image Folder: Tailored for training restoration diffusion models. It contains only Ground Truth (GT) and Low-Quality (LQ) images (i.e., clear and degraded images) without text annotations.
• MTJ_OCR Data: Contains the original scanned images along with their corresponding annotation files.

Dataset Details

Dataset Description

This dataset aims to address the complex layout recognition challenges encountered during the digitization of early historical documents. Because archives from the last century feature extensive Traditional Chinese, old-style Japanese (including obsolete Kanji), complex statistical tables, and special seals, this dataset provides a valuable benchmark environment for Document AI research in the fields of Computer Vision (CV) and Digital Humanities.

Dataset Sources

• Repository: https://huggingface.co/datasets/corste1/MTJ-OCR

Uses

Direct Use

• OCR Training: Character-level or line-level recognition targeting Traditional Chinese, variant characters, and old-style Japanese in vintage printed materials.

Out-of-Scope Use

• This dataset must not be used to forge historical documents or maliciously tamper with information in the original scans.

Dataset Creation

Curation Rationale

Due to the unique layout styles of early East Asian literature (e.g., multi-column mixed layouts, vertical text, mixed text and graphics), paper aging, and cross-lingual characteristics, general-purpose modern OCR models often perform poorly. By manually annotating this large-scale dataset, we can significantly enhance machine understanding and digitization capabilities for complex early documents.

Source Data

Data Collection and Processing

• Source: Scanned from original investigation reports, statistical yearbooks, and official archival documents from the first half of the 20th century.

Who are the source data producers?

The original data originated from various institutional publications and archives produced between the early and mid-to-late 20th century (approx. 1940s-1980s).

Bias, Risks, and Limitations

• Image Quality Limitations: Due to the printing techniques of the time and long-term storage conditions, some early archives suffer from physical defects such as ink bleed-through, blurred text, or torn paper. These factors increase the recognition difficulty for computer vision models.
• Linguistic and Textual Features: The text within the dataset reflects the specific linguistic habits and typesetting norms of its era. Models trained on this dataset may require fine-tuning when processing modern, standard text.

Citation

If you use this dataset in your research, please cite it as follows:

BibTeX:

@dataset{historical_document_ocr_2026,
  author = {YONG ZHENG},
  title = {Early 20th Century Historical Document OCR Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {[https://huggingface.co/datasets/corste1/MTJ-OCR](https://huggingface.co/datasets/corste1/MTJ-OCR)}
}