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
parisaocr epub book.pdf --out out                 # out/book.epub (experimental)
parisaocr epub book.pdf --out out --gemini        # the structure decided by Gemini (API key)
parisaocr epub book.pdf --out out --review        # then mark chapters by hand in the browser

Inputs are images, directories of images, or a PDF (scanned pages are extracted losslessly — a scan set inside a larger page, or under a site's name typed over it, is drawn from the embedded image where the page places it, without the stamp — and 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, and --pdf-text replace puts ours in place of the text of every scan page. A text layer of glyph codes in another script than the book's (a broken OCR, found in many shared scans) is replaced in any case. A site's name stamped as text over every page is removed, from the page images as well as from this PDF.
• 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.

E-books (EPUB), experimental

parisaocr epub book.pdf --out out --title "…" --author "…"

turns a scanned Persian book into an EPUB 3 e-book, out/book.epub, that reflows on phones and e-readers. Besides reading the pages, it works out the book from the page layout and the printed table of contents, helped by small line-role models (gradient-boosted trees over layout features, bundled; see MODEL_CARD.md). By default no language model is used and nothing leaves your machine; with a Gemini API key, --gemini (below) does much better. What it builds:

• chapters, with their titles and the book's table of contents; section headings, paragraphs joined across pages, block quotes;
• footnotes as popup notes linked to their markers (numbered per page or through the book, or with asterisks);
• verse: two-hemistich lines, and poems set line by line with their stanzas;
• figures and tables as images cut from the page (--tables html for HTML tables, but OCR of table numbers is not reliable enough to trust);
• the printed page numbers as the e-book's page list; pages put in book order by their printed numbers (reversed runs, duplicate scans and missing pages are found and reported);
• the book's title, author, publisher and ISBN from --title, --author etc. or a --meta JSON file; with --gemini they, and the translators and editors, are read from the cover, title and copyright pages.

Footnote numbers the OCR lost or cut are re-read from the page image (PARISAOCR_NOTENUM=0 turns this off). out/book.report.md lists what was decided and where to look first: pages read with low confidence, notes whose marker was not found, contents entries not found as chapters. The OCR is kept in out/book.work, so a second run takes seconds. Set EPUBCHECK_JAR to an epubcheck jar to have the result validated; --roles none uses the layout rules without the line-role models.

With a Gemini API key, --gemini gives the best structure. Every page image goes to Google's Gemini with its OCR lines numbered, and Gemini says what each line is: a chapter or section heading, a footnote or endnote, a running header, a byline, and so on. It then decides the book's outline from all the headings found and reads the title, authors, translators and publisher from the cover, title and copyright pages. On 10 test books that played no part in the development (measured with 0.4.0):

Most chapters it misses are found as section headings instead: books where the line between a chapter and a section is a judgement call. Put the key in GEMINI_API_KEY or ~/.config/gemini/api_key. Gemini 3.8 Flash costs about $0.004 to $0.005 a page (more for dense pages), $1.20 to $1.50 for a 300-page book; blank pages are not sent. Before sending anything, --gemini prints what this book should cost (within about 6% on the books we measured), and --gemini-estimate prints only that, after reading the pages locally. The answers are kept in out/book.work/gemini-…, so later runs and --review rebuilds cost nothing. The page images and their OCR text are sent to Google; without --gemini nothing leaves your machine.

Cheaper models through OpenRouter. --gemini --gemini-model openrouter:VENDOR/MODEL asks the same questions to an open model served by OpenRouter (key in OPENROUTER_API_KEY or ~/.config/openrouter/api_key). Only providers that keep no data, take none for training and enforce the answer's format are used. On two of our development books with many footnotes (326 pages, 64 footnotes checked):

¹ found with the right number and linked after the right word. ² with one provider for every page, PARISAOCR_OPENROUTER_PROVIDER='{"order": ["deepinfra/bf16"]}'; by default OpenRouter spreads pages over several providers, and their answers differ. Two books are too few to rank models closely; the Gemini figures above for 10 test books are the measured ones. PARISAOCR_OPENROUTER_PROVIDER is OpenRouter's provider object as JSON: it can name providers or exclude some, and add to the privacy settings, not weaken them. The cost estimate is calibrated on Gemini and only a rough guide for these models. Gemma 4 31B is also free from Google with a Gemini key (--gemini --gemini-model gemma-4-31b-it --gemini-jobs 2), but limited to 16,000 input tokens a minute: about 6 pages a minute, 50 minutes for a 300-page book; with more jobs Google refuses requests and the run stops after a few dozen pages (running it again continues).

Three minutes of your own marking make the chapters right whatever the scan, with or without Gemini. --review opens a local page (127.0.0.1, nothing leaves your machine) with every page as a thumbnail in reading order and the converter's decisions on them: parts, chapters and sections with their titles, figure pages. Click to add or remove an opening, fix a title (the OCR of the page's first lines is offered), set its level, then Rebuild. The marks are saved as out/book.marks.json, used by every later run of the book, and the converter's own decisions are kept beside them for comparison.

This is experimental: it was developed on 23 books (novels, a poetry anthology, histories with many footnotes, collections, a play, a book exported from Word) and measured on 10 others, and will meet layouts it gets wrong. Typical errors: a chapter title with an OCR error or cut short, a footnote linked at the end of its page instead of at its marker, an unusual heading taken for a paragraph. Multi-column pages (magazines) and dictionaries are not supported. Example pages of books it gets wrong are very welcome in the issues.

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.

کتاب الکترونیکی (EPUB، آزمایشی):

parisaocr epub book.pdf --out out --title "…" --author "…"
parisaocr epub book.pdf --out out --gemini      # ساختار کتاب با Gemini
parisaocr epub book.pdf --out out --review      # سپس علامت‌گذاری فصل‌ها در مرورگر

از کتاب اسکن‌شده یک کتاب EPUB می‌سازد که روی گوشی و کتاب‌خوان خوانده می‌شود: فصل‌ها و فهرست، پاراگراف‌ها، پانویس‌ها (با پیوند به شمارهٔ پانویس در متن)، شعر، تصویرها و جدول‌ها، و شمارهٔ صفحه‌های چاپی. گزارشی هم (out/book.report.md) می‌نویسد که نشان می‌دهد کجا را باید بررسی کرد. این قابلیت آزمایشی است؛ اگر کتابی درست تبدیل نشد، لطفاً در بخش Issues خبر بدهید.

اگر کلید API برای Gemini دارید، با --gemini ساختار کتاب بسیار بهتر درمی‌آید: تصویر هر صفحه همراه با سطرهای شماره‌خوردهٔ اوسی‌آر به Gemini گوگل فرستاده می‌شود تا بگوید هر سطر چیست (عنوان فصل یا زیربخش، پانویس، سرصفحه و …)؛ سپس ساختار کل کتاب و مشخصات آن (عنوان، نویسنده، مترجم، ناشر) را از روی همهٔ عنوان‌ها و صفحه‌های جلد و شناسنامه تعیین می‌کند. روی ۱۰ کتاب آزمون که در ساختن ابزار به کار نرفته بودند، ۸۶٪ فصل‌ها و ۸۹٪ پانویس‌ها را پیدا کرد، در برابر ۴۸٪ و ۶۵٪ بدون آن. کلید را در GEMINI_API_KEY یا در فایل ~/.config/gemini/api_key بگذارید. هزینه با Gemini 3.8 Flash برای هر صفحه حدود ۰٫۰۰۴ تا ۰٫۰۰۵ دلار است (۱٫۲ تا ۱٫۵ دلار برای کتابی ۳۰۰ صفحه‌ای). پیش از ارسال، هزینهٔ تقریبی همان کتاب چاپ می‌شود، و با --gemini-estimate فقط همین برآورد را می‌بینید و چیزی فرستاده نمی‌شود. پاسخ‌ها نگه داشته می‌شوند، پس اجرای دوباره هزینه‌ای ندارد. با این گزینه تصویر صفحه‌ها و متن اوسی‌آر به گوگل فرستاده می‌شود؛ بدون آن هیچ چیز از رایانهٔ شما بیرون نمی‌رود.

مدل‌های ارزان‌تر از راه OpenRouter: با --gemini --gemini-model openrouter:qwen/qwen3.8-27b (حدود ۰٫۵ دلار برای کتابی ۳۰۰ صفحه‌ای) یا --gemini --gemini-model openrouter:google/gemma-4-31b-it (حدود ۰٫۱۵ دلار) همین پرسش‌ها از یک مدل باز پرسیده می‌شود. کلید را در OPENROUTER_API_KEY یا فایل ~/.config/openrouter/api_key بگذارید. فقط ارائه‌دهنده‌هایی به کار می‌روند که داده‌ای نگه نمی‌دارند و برای آموزش استفاده نمی‌کنند. روی دو کتاب پرپانویس، هر دو مدل همهٔ فصل‌ها را پیدا کردند و ۷۹ تا ۸۸ درصد پانویس‌ها را درست پیوند دادند (Gemini: ۹۲ درصد).

با --review صفحه‌ای در مرورگر خودتان باز می‌شود که همهٔ صفحه‌های کتاب را کوچک و به ترتیب نشان می‌دهد. آغاز بخش‌ها، فصل‌ها و زیربخش‌ها و عنوانشان را درست کنید و کتاب را دوباره بسازید؛ این علامت‌ها ذخیره می‌شوند و در اجراهای بعدی هم به کار می‌روند.

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.