ParisaOCR (پریسا اوسی‌آر)

Open-source OCR for printed Persian. It reads scanned book pages about as well as Google's Gemini Flash models, runs on your own machine (CPU or GPU, no internet needed), and costs a few cents per thousand pages to run.

Status: preview. It works well on printed Persian books and documents; see Known limitations. Feedback with example pages is very welcome.

فارسی · Model on Hugging Face · PyPI

Install

pip install parisaocr

Python 3.10 to 3.13. The two models (text-line detector and recognizer, 22 MB) are included. PyTorch is installed as a dependency of Kraken; a GPU is used when one is available, otherwise the CPU (about 5 to 10 seconds per page). For PDF input, Poppler's command-line tools must be installed (poppler-utils on Debian and Ubuntu, poppler on Arch, Homebrew and conda-forge).

Use

parisaocr ocr page.png                        # print the text
parisaocr ocr scans/ --out out                # out/txt/*.txt and out/hocr/*.hocr
parisaocr ocr book.pdf --out out --format txt,hocr,jsonl
parisaocr ocr book.pdf --out out --first 10 --last 20
parisaocr ocr book.pdf --out out --format pdf     # out/pdf/book.pdf: searchable

Inputs are images, directories of images, or a PDF (scanned pages are extracted losslessly, born-digital pages rendered). Outputs: plain text (one line per printed line, columns right to left), hOCR with line and word boxes on the original page, and JSONL with text, boxes and confidences.

Searchable PDFs

--format pdf writes PDFs you can search, select and copy text in, with the page images exactly as they were:

• PDF input: out/pdf/NAME.pdf is the original file with an invisible text layer over each page that was read. The scans are not re-encoded; page rotation is respected. Pages that already have a text layer (born-digital pages, or scans someone OCRed before) are left as they are; --pdf-text add adds ours to them too, e.g. over a poor earlier OCR layer.
• Image input: one PDF per image in out/pdf/, or all of them in one file with --merge-pdf NAME. The page size follows the image's dpi.

The text layer is stored the way word processors store Persian text (one run per line in visual order), so PDF viewers find and copy it in reading order. Tested with Poppler (pdftotext, used by Okular and Evince) and pdf.js (Firefox): words and their order come out right; pdftotext sometimes moves punctuation at the edge of a word, and pdf.js drops half-spaces when copying. Tools that do not reorder right-to-left text (such as pdfminer) return each line reversed, as they do for any Persian PDF.

From Python:

from parisaocr import OCR

ocr = OCR()                         # loads the models once
print(ocr.text("page.png"))
for line in ocr.lines("page.png"):  # reading order
    print(line.bbox, line.conf, line.text)

Accuracy

Measured on material the model never saw in training.

¹ Our earlier Persian Tesseract model (tessdata_contrib), reading the lines found by the Surya detector.

On the benchmark the gap to Gemini comes mostly from very small, low-resolution scans (lines under about 16 pixels high), where Gemini reads about 78% of the words and ParisaOCR 74%. Many benchmark pages are hard even for a human reader, and its page-level character error (about 40% for every reader) mostly measures reading order on multi-column pages.

The verified test set is small (750 lines from 11 books), so treat the numbers as a guide rather than a precise ranking. Details in MODEL_CARD.md.

Known limitations

• Printed Persian only. Handwriting, Nastaliq calligraphy, manuscripts and Urdu are not supported.
• Very small text (lines under about 14 pixels high, i.e. low-resolution scans) loses accuracy; ParisaOCR prints a warning for such pages. Scan at 300 dpi when you can.
• Layout is simple. Columns are read right to left and top to bottom; tables, captions and complex magazine layouts may come out in the wrong order. Text rotated sideways is not read.
• Small ornaments (a box, a dingbat) are sometimes read as a letter.
• Spacing conventions. The model follows standard Persian spelling for half-spaces (ZWNJ) and sometimes differs from the printed spacing.
• Diacritics (vowel marks) are read less reliably than letters.

How it works

1. Lines. The PP-OCRv6 text detector (PaddleOCR, Apache-2.0) finds the text lines, run with ONNX Runtime.
2. Crops. Each line is cut out with a margin that keeps the dots above and below the letters; neighbouring lines inside the margin are painted over.
3. Reading. A Kraken recognition model (PP-OCR-style architecture, CTC) trained for Persian print reads each line. ParisaOCR turns the model's visual-order output into reading order itself, keeping the half-space (ZWNJ), and with each line's own direction, so English lines and footnotes come out right.
4. Order. Lines are grouped into columns, read right to left.

The model was trained on about 83,000 lines from about 200 scanned Persian books, labelled automatically with Gemini, plus about 90,000 synthetic lines (Persian and English text, special characters, low-resolution copies). See MODEL_CARD.md, and the modules in docs/architecture.md.

Other models and detectors

--model accepts another Kraken model (kraken:model.safetensors) or a Tesseract model (TESSDATA_DIR:LANG); --detector accepts other PP-OCR sizes (ppocr:v6-medium), Kraken's segmenter (kraken), or Surya (surya, after pip install parisaocr[surya]; note that Surya's weights are free only for research, personal use and small companies). parisaocr cut cuts the lines of scanned books for building training data.

Feedback

Please open an issue with the page (only if you are allowed to share it), the command you ran, and what came out wrong.

License

Apache License 2.0, for the code and the models. The bundled PP-OCRv6 detector is PaddleOCR's (Apache-2.0); see NOTICE.

فارسی

پریسا اوسی‌آر یک اوسی‌آر متن‌باز برای متن چاپی فارسی است. صفحه‌های اسکن‌شدهٔ کتاب را تقریباً هم‌اندازهٔ مدل‌های Gemini Flash گوگل می‌خواند، روی رایانهٔ خودتان اجرا می‌شود (بدون نیاز به اینترنت، با CPU یا GPU) و هزینهٔ اجرایش برای هر هزار صفحه چند سنت است.

وضعیت: پیش‌نمایش. روی کتاب‌ها و اسناد چاپی فارسی خوب کار می‌کند. محدودیت‌ها را پایین‌تر ببینید. بازخورد همراه با صفحهٔ نمونه بسیار ارزشمند است.

نصب:

pip install parisaocr

استفاده:

parisaocr ocr page.png                      # چاپ متن
parisaocr ocr book.pdf --out out            # out/txt و out/hocr
parisaocr ocr book.pdf --out out --format pdf   # out/pdf/book.pdf: پی‌دی‌اف جست‌وجوپذیر

با --format pdf خروجی یک پی‌دی‌اف جست‌وجوپذیر است: تصویر صفحه‌ها همان است که بود و یک لایهٔ متن نامرئی روی آن قرار می‌گیرد، تا بتوانید در آن جست‌وجو کنید و متن را انتخاب و کپی کنید.

محدودیت‌ها:

• فقط متن چاپی فارسی. دست‌نوشته، خط نستعلیق، نسخهٔ خطی و اردو پشتیبانی نمی‌شوند.
• متن خیلی ریز (اسکن با وضوح پایین) دقت کمتری دارد. در صورت امکان با ۳۰۰ dpi اسکن کنید.
• چیدمان صفحه ساده خوانده می‌شود. جدول‌ها و صفحه‌های مجله‌ای پیچیده ممکن است ترتیب درستی نداشته باشند.
• نیم‌فاصله‌ها طبق املای استاندارد نوشته می‌شوند و گاهی با چاپ کتاب فرق دارند.
• اِعراب‌ها کمتر از حروف دقیق خوانده می‌شوند.

بازخورد: لطفاً در بخش Issues صفحه (فقط اگر اجازهٔ انتشارش را دارید)، فرمانی که اجرا کردید و خطای خروجی را بفرستید.

مجوز: Apache 2.0، برای کد و مدل‌ها.