Sigurdur/isl_synthetic_ocr

Icelandic Synthetic OCR Dataset

Dataset Description

This dataset contains synthetically generated images of Icelandic text for Optical Character Recognition (OCR) tasks. It was created to address the lack of publicly available, large-scale OCR datasets for the Icelandic language. The images are generated from a corpus of Icelandic text, rendered with various settings to create a diverse set of training examples.

The key feature of this dataset is its focus on Icelandic, including all special characters such as ð, þ, æ, ö, á, é, í, ó, ú.

Project Status: Work in Progress

This is an ongoing project. The dataset is currently under development, and future versions may include more variety in:

• Fonts and text styles
• Backgrounds and textures
• Real-world augmentations (e.g., noise, blur, perspective distortion)

Feedback and contributions are welcome.

How to Use

You can load the dataset from the Hugging Face Hub like this:

from datasets import load_dataset

dataset = load_dataset("Sigurdur/isl_synthetic_ocr")

Dataset Structure

Data Instances

A typical data instance consists of an image and its corresponding ground truth text.

{
  'image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=512x512>,
  'text': 'Íslenskur texti'
}

Data Fields

• image: A PIL image object containing the rendered text.
• text: A string containing the ground truth text corresponding to the image.

Data Splits

This dataset does not have pre-defined train/validation/test splits. It is provided as a single collection of generated examples. Users are encouraged to create their own splits as needed for their specific use case. This is subject for change later in the development process.

Dataset Creation

Curation Rationale

The primary motivation for creating this dataset was the scarcity of dedicated OCR training data for Icelandic. By synthetically generating images from a large text corpus, we can create a virtually unlimited amount of labeled data, which is crucial for training robust deep learning models for OCR.

Source Data

The text used for generating the images is sourced from the arnastofnun/IGC-2024 dataset, specifically the wiki subset, which contains a large collection of Icelandic text. I chose the wiki subset because it contains many domain specific words.

Generation Process

The images were generated using a Python script leveraging the Pillow library. The script takes text from the source corpus and renders it onto a blank image. The generation process is configurable, allowing for variations in:

• Image dimensions (width, height, dpi)
• Font type, size, and color
• Text alignment (horizontal and vertical)
• Background color

The code used for generation is available in the dataset repository.

Limitations and Bias

• Synthetic Nature: The images are "clean" and do not feature real-world distortions like noise, blur, perspective shifts, or varied lighting conditions. Fine-tuning on real-world data may be necessary for robust performance.
• Font Variety: The dataset was generated using a limited set of fonts (e.g., Arial). OCR models trained on this data may not generalize well to documents with different or more stylized fonts.
• Text Corpus Bias: The dataset inherits any biases present in the source arnastofnun/IGC-2024. The content is primarily from news and web text, and may not represent all styles of written Icelandic.

Citation

If you use this dataset in your research, please consider citing it:

@misc{isl_synthetic_ocr,
  author    = {Sigurdur Haukur Birgisson},
  title     = {Icelandic Synthetic OCR Dataset},
  year      = {2025},
  publisher = {Hugging Face},
  journal   = {Hugging Face repository},
  howpublished = {\url{https://huggingface.co/datasets/Sigurdur/isl_synthetic_ocr}}
}