ParisaOCR recognition model 0.1

A text-line recognizer for printed Persian, the reading model of ParisaOCR. It reads one cropped line image at a time; ParisaOCR adds line detection (PP-OCRv6), cropping and reading order, and is the intended way to use it:

pip install parisaocr
parisaocr ocr page.png

Model

• Architecture: Kraken 7.1's ppocrv6 recognizer (PP-OCR-style convolutional backbone with an SVTR neck and a CTC head), variant small, about 3 million parameters; input lines are scaled to 96 pixels high.
• Output: Persian and Latin letters, Persian and Latin digits, punctuation, common symbols (٪ % ٫ ٬ × ÷ ° ± $ € £ and more), diacritics and the zero-width non-joiner (ZWNJ, half-space); 182 classes.
• File: parisaocr-fa-0.1.safetensors, 12 MB.

Use it through ParisaOCR. Plain kraken ocr gives worse results with this model, for two reasons we reported to Kraken: PP-OCR models must see inputs padded to the training width (mittagessen/kraken#810, otherwise characters are invented at the start of Persian lines), and Kraken's bidi reordering deletes the ZWNJ (mittagessen/kraken#809). The model was trained on labels put into visual order by ParisaOCR's own reordering (parisaocr/bidi.py, with each line's own base direction), which ParisaOCR also applies to its output.

Training data

• Books. About 200 scanned Persian books, mostly 20th-century and modern print, from 60 to over 600 dpi, from a public online collection whose copyright status varies and is unknown to us. Twelve pages per book were used. The scans, line images and labels are not redistributed; only the trained model is.
• Labels. Each real line was transcribed once, automatically, by Google Gemini (3.8 Flash for most books, 3.7 Flash for 40), reading line images in sheets of eight, in standard Persian spelling. Labels were not checked by people. On a human-verified test set the teacher's character error, ignoring spaces, is 0.25-0.36%, a rough measure of the label noise.
• Synthetic lines were rendered with Tesseract's text2image and degraded to look like scans (blur, noise, resolution loss, JPEG, binarization). Wikipedia text is licensed CC BY-SA 4.0.
• Training. From scratch for 12 epochs on a first set (round one, about 100 books), fine-tuned for 4 epochs with the ZWNJ kept, then fine-tuned for 4 epochs on the set above (round two). One RTX 4070 Ti, about 15 GPU hours in total. Validation used 6 whole books held out of training.

Evaluation

All test material is from books and fonts that were not used for training.

The verified lines keep the spacing of the print; the model and Gemini write standard spelling, so measures that count spaces penalize both slightly. The verified set is small, so differences of a few tenths of a percent are within noise. The three synthetic line sets are rendered, not scanned.

Limitations

• Printed Persian only: not handwriting, Nastaliq, manuscripts or Urdu.
• Accuracy drops on very small text (lines under about 14 pixels high).
• Diacritics and rare symbols are less reliable than letters.
• Half-spaces follow standard spelling, which can differ from the print.
• A label noise floor from the automatic teacher.

Line-role models (EPUB structure)

parisaocr epub also bundles five small models (parisaocr/models/roles/, 3.7 MB) that give each OCR line a role, which the converter's layout rules consult when they work out a book's chapters and footnotes.

• Model. Gradient-boosted decision trees (scikit-learn HistGradientBoostingClassifier, 300 iterations, 31 leaves; exported as plain node arrays and evaluated with numpy) over 46 features of a line: its box and size relative to the page and the book, position among the page's lines, gaps to its neighbours, text shape (length, digits, Latin letters, a leading number, a label word such as فصل), OCR confidence, how often its text repeats at the top of pages, ink density, printed rules above it, and the small type below it. Tasks: the line role (13 classes: running header, page number, heading, body, quote, verse, footnote, endnote, caption, table, figure, contents, other), footnote or not, heading or not, whether a footnote line starts a note, and the heading level (1-3).
• Training data. 121,222 lines on 4,018 sampled pages of 85 scanned Persian books (the books of the recognizer's collection, OCRed by ParisaOCR; the pages were chosen to include footnotes and chapter openings). The roles were labelled automatically by Google Gemini 3.8 Flash from the page image and the OCR lines; on 518 pages with a human-made structure truth the labels found 97% of chapter openings and footnote lines with precision 92% / recall 95%. The pages, lines and labels are not redistributed.
• Evaluation. Through the whole converter on 8 held-out books with a human-made truth (chapters, footnotes, section headings), against the layout rules alone: chapter recall 57% → 64% (precision 60% → 67%), footnotes found 44% → 54% (pooled), footnotes linked at their marker 62% → 76%, section headings 46% → 63%. Leave-one-book-out on the 15 development books: footnote-line F1 85%, chapter-opening F1 86%.
• Limitations. Trained on books of the same collection as the recognizer (20th-century and modern print, Persian); titles set in display type that the OCR does not read, and footnotes without a separator rule, remain the main losses.

License

Apache License 2.0.