Arabic OCR on Windows (Tesseract + OCRmyPDF)

This repo provides a fast, high‑accuracy, fully local pipeline to extract Arabic text from PDFs on Windows using:

• Tesseract 5 with the Arabic model from tessdata_best (highest free accuracy).
• OCRmyPDF as the smart wrapper that auto‑rotates, deskews, cleans pages, and produces a searchable PDF plus a sidecar TXT file.

ℹ️ Reality check: Zero OCR errors is not realistic, especially with Arabic (fonts, scan quality, diacritics). The settings here balance accuracy, speed, and reasonable file sizes for long PDFs.

---

✅ Outputs

• Searchable PDF: *_OCR.pdf
• Plain text sidecar: *_OCR.txt — strongly recommended for search/indexing and to bypass some RTL quirks in certain PDF viewers.

---

🧩 Requirements (Windows)

System (install via Chocolatey):

• Python 3.10+
• Tesseract OCR
• Ghostscript
• (Optional) pngquant (better compression for color images)

Python:

• ocrmypdf (installed via pip)

Ensure TESSDATA_PREFIX points to your Tesseract tessdata directory. You must install Arabic model (ara) from tessdata_best.

---

⚡️ Quick Install (PowerShell — Run as Administrator)

If you don’t have Chocolatey, install it from its official site first. Then run:

# 1) System tools
choco install -y python3 tesseract ghostscript pngquant

# 2) Python packages
pip install --upgrade pip
pip install -r requirements.txt

# 3) Arabic model from tessdata_best
Invoke-WebRequest `
  https://github.com/tesseract-ocr/tessdata_best/raw/main/ara.traineddata `
  -OutFile "C:\Program Files\Tesseract-OCR\tessdata\ara.traineddata"

# 4) Make sure Tesseract knows where models live
[Environment]::SetEnvironmentVariable('TESSDATA_PREFIX', 'C:\Program Files\Tesseract-OCR\tessdata', 'Machine')
$env:TESSDATA_PREFIX='C:\Program Files\Tesseract-OCR\tessdata'

# 5) Sanity checks
tesseract --version
tesseract --list-langs     # "ara" should be listed
ocrmypdf --version

If you see permission issues, reopen PowerShell as Administrator. If ara is missing, confirm ara.traineddata is in the correct tessdata directory and that TESSDATA_PREFIX is set.

---

🗂️ Suggested Project Layout

OCR/
├─ app.py           # main script
├─ requirements.txt # Python deps
└─ README.md

---

▶️ How to Run (Python script)

Place your PDF somewhere (example): C:\Users\waok\Downloads\ARABIC PDF PATH .pdf

From the project folder:

python app.py "C:\Users\waok\Downloads\ARABIC PDF PATH .pdf" "C:\Users\waok\Downloads\ARABIC PDF PATH _OCR.pdf"

If you omit arguments, app.py uses a default input path inside the script. The script also writes a sidecar TXT next to your output PDF: ..._OCR.txt.

Direct CLI (no Python)

ocrmypdf -l ara --jobs 8 `
  --rotate-pages --deskew --clean --remove-background `
  --optimize 3 --output-type pdf `
  --pdf-renderer hocr `
  --sidecar "C:\Users\waok\Downloads\ARABIC PDF PATH _OCR.txt" `
  "C:\Users\waok\Downloads\ARABIC PDF PATH .pdf" `
  "C:\Users\waok\Downloads\ARABIC PDF PATH _OCR.pdf"

If you see “page already has text” and still want to force OCR, add --force-ocr.

---

🧠 What the Script Does (Key Settings)

• Language: ara (add +eng if your content is mixed).
• Preprocessing: Auto‑rotate, deskew, and background clean before OCR for better accuracy.
• Tesseract engine: LSTM‑only (most accurate) with automatic page segmentation (PSM 3) — good default for diverse document layouts.
• PDF rendering: pdf_renderer="hocr" works better with RTL languages in many viewers.
• Output type: output_type="pdf" (smaller than PDF/A).
• Sidecar: *_OCR.txt is reliable for search/indexing and downstream NLP.
• Parallelism: jobs leverages CPU cores to speed up long PDFs.

---

✍️ Suggested app.py

Tuned for highest free accuracy with reasonable output size on Windows.

import os, sys, pathlib
import ocrmypdf

def main(input_path: str, output_path: str | None = None):
    # Ensure Tesseract models path (tessdata_best) is known
    os.environ.setdefault("TESSDATA_PREFIX", r"C:\Program Files\Tesseract-OCR\tessdata")

    p_in = pathlib.Path(input_path)
    if output_path is None:
        output_path = str(p_in.with_name(p_in.stem + "_OCR.pdf"))
    sidecar = str(pathlib.Path(output_path).with_suffix(".txt"))

    # NOTE: If your PDF is truly 300 dpi or higher, remove oversample=300
    ocrmypdf.ocr(
        str(p_in),
        output_path,
        language="ara",                 # add +eng if you have mixed Arabic/English
        rotate_pages=True,
        deskew=True,
        clean=True,
        remove_background=True,
        optimize=3,
        output_type="pdf",
        pdf_renderer="hocr",
        tesseract_oem=1,                # LSTM only
        tesseract_pagesegmode=3,        # automatic page segmentation
        oversample=300,                 # drop this if input is already >=300 dpi
        sidecar=sidecar,
        jobs=max(1, (os.cpu_count() or 8) - 1),
    )
    print(f"✅ Done:\nPDF : {output_path}\nTXT : {sidecar}")

if __name__ == "__main__":
    default_input = r"C:\Users\waok\Downloads\ARABIC PDF PATH .pdf"
    in_path = sys.argv[1] if len(sys.argv) > 1 else default_input
    out_path = sys.argv[2] if len(sys.argv) > 2 else None
    try:
        main(in_path, out_path)
    except Exception as exc:
        print("Failed to run ocrmypdf:", exc)
        sys.exit(2)

---

🔧 Quality & Speed Tuning

• PSM (page segmentation):
  • tesseract_pagesegmode=3 is a strong default for varied layouts.
  • =6 is great for uniform text blocks (paragraph pages).
• OEM (engine):
  • tesseract_oem=1 (LSTM‑only) is typically the most accurate.
• Languages:
  • Use ara+eng only if needed; adding more languages can sometimes degrade accuracy.
• Oversample:
  • Keep oversample=300 only when scans are below 300 dpi. Otherwise remove it to reduce output size.

---

🧰 Troubleshooting

• WinError 2 / “command not found”: Add Tesseract and Ghostscript to your PATH or reopen PowerShell as Admin after install.
• “Tesseract couldn’t load any languages”: Confirm TESSDATA_PREFIX and that ara.traineddata exists in the tessdata folder.
• Large output size:
  • Prefer output_type="pdf" (smaller than default PDF/A).
  • Use optimize=3.
  • Remove oversample if scans are already 300+ dpi.
  • jbig2 is commonly missing on Windows; it’s safe to ignore (only affects B/W compression).
• Warning lots of diacritics: Informational; common with Arabic. Improve source quality and keep effective dpi around 300.
• RTL quirks in PDF viewers: Some viewers struggle with Arabic text selection/search. Rely on the sidecar TXT or try another viewer.

---

🧽 (Optional) Export TXT without diacritics

import re, io
in_txt  = r"C:\path\to\file_OCR.txt"
out_txt = r"C:\path\to\file_OCR_no_diac.txt"

with io.open(in_txt, "r", encoding="utf-8") as f:
    txt = f.read()

# Remove common Arabic diacritics
txt_no_diac = re.sub(r"[\u064B-\u065F\u0670\u06D6-\u06ED]", "", txt)

with io.open(out_txt, "w", encoding="utf-8") as f:
    f.write(txt_no_diac)

print("Saved:", out_txt)

---

🧪 Quick Test

# Replace the path with your PDF
python app.py "C:\Users\...\ARABIC PDF PATH .pdf"
# Check outputs:
#   ...\ARABIC PDF PATH _OCR.pdf
#   ...\ARABIC PDF PATH _OCR.txt

---

📄 License

Free to use internally. Respect the upstream licenses (Tesseract, OCRmyPDF).

---

💬 Need a one‑click .bat?

I can include a Windows Batch (.bat) that asks for a PDF path and runs the same settings automatically (and writes the TXT next to it). Just say the word.