What this does
Upload an image or PDF of an Icelandic document and get the text back, line by line. Each line is segmented independently. The output includes a confidence score (0–1), a boundary polygon, and a baseline, all overlaid on the original.
The engine is Kraken. The recognition model is catmus-print-fondue-ft, fine-tuned on Icelandic text. It handles Þ/þ, Ð/ð, Æ/æ and accented vowels that generic OCR engines get wrong.
The base model is CATMuS-Print (Gabay & Clérice, 2024, Zenodo, 10.5281/zenodo.10592716).
The model is small
This model was fine-tuned on ∼100 Icelandic documents. It works well on clean, modern text. Historical fonts, low resolution, and unusual layouts will trip it up. That is why every line comes with a confidence score.
You can make it better. After each run, hit the thumbs up or thumbs down to rate the result. If a document looks like it could help, turn on auto-donate in your account settings. Your uploaded image gets folded into future training, with no account info attached.
Why this exists
Seven million pages of historical Icelandic newspapers and periodicals (Tímarit.is) are locked in archives. No off-the-shelf OCR works on our language. Standard tools like Tesseract and Google Vision miss the special characters that matter for Icelandic.
I built this as part of a broader research effort into synthetic-data pipelines for low-resource languages. The goal is the same as it has always been: make technology work for a language the world overlooks.
I am Sigurdur Haukur. I am from Iceland, finishing my AI degree at University of Groningen, competing in cybersecurity and sailing for Iceland. The broader project lives on sigurdurhaukur.com.
Quotas
One PDF per day as a guest. Ten per day if you sign up and verify your email. Need more for research, archival work, commercial use, or integration? Email contact@sigurdurhaukur.com.
Icelandic OCR, by Sigurdur Haukur