freococo/myanmar_typeset_dictionary_OCR

Myanmar Typeset Dictionary OCR Dataset

This is a synthetically generated, realistically formatted dataset modeling a Myanmar-Myanmar dictionary. It is designed for training and validating OCR models, Document Layout Analysis (DLA) pipelines, and structural key-value extraction models.

The dataset contains a highly diverse set of pages containing multiple column flows, tabular glossaries, running headers/footers, realistic backgrounds, and dynamic typography (four fonts paired per page for Word, POS, Pronunciation, and Definition).

• —Hugging Face Repository: freococo/myanmar_typeset_dictionary_OCR
• —Total Pages: 886
• —Resolution: 150 DPI (Crisp, clean baseline ideal for downstream data augmentations like blurs and rotations)
• —Language: Myanmar (Burmese)

Repository Structure

The repository follows Hugging Face's native ImageFolder structure:

myanmar_typeset_dictionary_OCR/
├── images/
│   ├── page_0001.png
│   ├── page_0002.png
│   └── ... (all 886 images)
├── metadata.jsonl
└── README.md

Metadata Schema (metadata.jsonl)

Each JSON line in metadata.jsonl represents a single page with the following structure:

• —`file_name`: Path to the image relative to the repository root.
• —`layout_style`: The document layout template used (1-column-portrait, 2-column-portrait, 3-column-portrait, 3-column-landscape, 4-column-landscape, table-borders, table-noborders, table-striped, table-compact, table-striped-compact).
• —`page_number`: Page index.
• —`width` / `height`: Rendered dimensions.
• —`page_size_name`: Standard page dimension (A4, A5, B5, US Trade, Letter).
• —`fonts_used`: Dictionary detailing the specific fonts rendered for the word, pos (part of speech), sound (pronunciation), and meaning (definition) fields.
• —`header` / `footer`: Bounding box and string transcription.
• —`columns`: Bounding box arrays for each text column.
• —`table_bbox`: Bounding box of the outer table element (Style 6 to 10).
• —`entries`: List of individual dictionary entries with their bounding boxes, clean labels, and exact visual transcripts (including visual delimiters like () and []).