fgaim/GLOCR-Tigrinya

GLOCR: GeezLab OCR Dataset

Overview

GLOCR is a Text Recognition (TR) and Optical Character Recognition (OCR) dataset for the Tigrinya language. The dataset contains a total of 661K image-label pairs from multiple data sources. In addition to the characters-only data, the major part of the dataset is a collection of multi-word text images with labels from three categories: News (from Haddas Ertra newspaper), the Bible, and random-trigrams of the 150k most common words in Tigrinya.

Dataset Summary

• Total samples: ~661K image-label pairs
• Total size: >1.3GB (tar.gz archives)
• DOI: 10.7910/DVN/RQTSD2

Dataset Subsets

Config	Description	Train	Validation	Test
news	Newspaper text-lines	200K	15K	15K
bible	Biblical text-lines	80K	10K	10K
top150k	Word trigrams	150K	15K	15K
characters	Single characters	120K	15K	15K
unsegmented	Full-page scans	506	-	-

Usage

Loading a specific subset

from datasets import load_dataset

# Load a specific subset, one of: news, bible, top150k, characters, unsegmented
news = load_dataset("fgaim/GLOCR-Tigrinya", "news")

# Access samples
sample = news["train"][0]
print(sample["text"])
sample["image"].show()

Loading a specific split

# Load a specific split of a subset
bible_test = load_dataset("fgaim/GLOCR-Tigrinya", "bible", split="test")

# Access samples
print(bible_test["text"][0])
bible_test["image"][0].show()

Loading all text-line data combined

# Load all text-line data combined
all_data = load_dataset("fgaim/GLOCR-Tigrinya", "all")

# Access samples
sample = all_data["train"][0]
print(sample["text"])
sample["image"].show()

Links

• Harvard Dataverse
• GitHub Repository

Citation

@data{gaim-2021-glocr,
  title     = {{GLOCR: GeezLab OCR Dataset}},
  author    = {Fitsum Gaim},
  year      = {2021},
  month     = {April},
  publisher = {Harvard Dataverse},
  version   = {1.0},
  doi       = {10.7910/DVN/RQTSD2},
  url       = {https://doi.org/10.7910/DVN/RQTSD2},
  dataverse = {https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/RQTSD2}
}

License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.