Nau OCR 1.0 (ނައޫ)

Reads Dhivehi (Thaana) text from scanned and born-digital PDFs and images, on an ordinary CPU, and says where every line and word is on the page. Pages that arrive sideways, upside down or tilted are put right before reading.

Update, 11 October 2026: a new line cut, and seal removal

The reader now cuts lines so that every mark stays with its own line. On single-spaced pages the fili above one line sit in the same rows as the bottom of the line above. The old line finder split two such lines at a single row, so a line's marks often ended up in the strip above or in a band of their own, and the recogniser saw consonants with nothing over them. The new cut gives each connected piece of ink to the line whose letters it belongs to. It also folds a band of marks back into its line and paints each neighbour's ink out of a strip. Lines that read with all their marks wrong ("collapsed fili") fell by half on the pages below. On the full 300-page independent benchmark, its maintainers measured round 11's median error falling from 1.3% to 0.7% with the new cut.

nau_ocr.py --remove-seal takes a coloured seal out of a page before its lines are found. It removes only a hollow, round, coloured ring 120–500 px across at 200 dpi, and the coloured ink inside it. It leaves the ring alone if type overlaps it, and never removes black ink. It is off unless asked for. The weights are unchanged; all figures below are with the new reader.

Update, 9 October 2026: new weights

The first weights published here (until 9 October 2026) read born-digital documents (Word and PDF exports) worse than the recogniser they were built from. They dropped full stops at the ends of lines, numbers at the start of list items, brackets, commas and some spaces. Their training lines were cut from the text layer with padding, so many line images showed marks their labels did not have. The model learned to leave out what it could see. The "printed documents" figure published with them was measured on lines made the same way, so the problem did not show.

The current weights are trained again from the original recogniser. Every born-digital training line is cut from the whole page by the reader's own line finder. Its label is the page's own text for exactly that strip, and a line is used only when all of its ink is covered by that text. Pages whose text layer is damaged are left out. All scores below are on whole pages or on lines labelled that way. The first weights stay in this repository's history, at commit 238ba8f.

Results

Whole born-digital pages, read as a user would read them, scored against each page's own text (character error rate over the page; a known line counts as found when a read line is within a quarter of its length):

The worst pages in both sets are dense tables, where some lines are missed.

Held-out lines (character error; lines read exactly in brackets):

Whole scanned pages, turned and tilted at random (218 pages from every decade, each given a random quarter turn and up to ±5° tilt). Score: the share of words read that are real Dhivehi words.

nau_ocr.py now keeps a quarter turn only when the recogniser reads the turned page clearly better. The ink-profile test alone turned some upright pages. Given each of four turns, it puts 116 of 120 born-digital pages and 154 of 160 scanned pages (1980s–2020s) the right way up.

What it does not read well yet

• Handwriting. Many 1960s–1980s documents are handwritten; the model has not been trained on handwriting.
• Arabic script and calligraphy (the 1930s–1940s records).
• Dense tables: lines are missed where cells crowd together.
• Text crossing table borders, and italic faces: read, but with more dropped letters than plain text.
• Output from old documents should be proofread before it is quoted.

Running it on CPU

pip install torch pymupdf pillow numpy
python nau_ocr.py document.pdf            # writes document.txt and document.json
python nau_ocr.py scan.png --pages 1-3
python nau_ocr.py sealed.pdf --remove-seal     # take a coloured seal out before reading

document.json holds, per page: the turn and deskew applied, the text, and every line and word with its box as fractions of the prepared page (origin top-left). Word boxes are placed from where the model read each character; line boxes are exact.

from nau_ocr import NauOCR, pages_of
ocr = NauOCR("recogniser.pt")
for n, page in pages_of(Path("document.pdf"), None):
    result = ocr.read(page)          # {"text", "lines": [{"text", "box", "words": [...]}], ...}

recogniser.pt holds the recogniser (checkpoint["model"], checkpoint["charset"]) and, beside it, the upside-down head (checkpoint["updown"]).

Licence and attribution

CC BY-NC-SA 4.0, with a commercial licence on request.

• Non-commercial use (research, archives, teaching, public-interest projects) is free. Credit Nau OCR, and share anything you build from it under the same licence.
• Commercial use needs a commercial licence from the Nau authors; its terms include publishing the improvements you make. Ask by opening a discussion titled "Commercial licence" on this repository.

See NOTICE.

Access requires accepting these terms in the form at the top of this page.

Related: Nau 1.0 (Dhivehi chat model) and Nau Embed 1.0 (Dhivehi search embeddings).