UTRNet PDF OCR — Coolify Container

End-to-end Urdu PDF OCR. Upload a PDF, get back a ZIP with one .txt file per page.

Pipeline: PDF → Page Images (200 DPI) → YOLOv8 line detection → UTRNet-Large recognition → per-page .txt files → ZIP

---

Files

utrnet-pdf/
├── Dockerfile          # Python 3.10 slim, poppler, CPU PyTorch 2.0.1
├── docker-compose.yml  # Persistent model volume, healthcheck
├── entrypoint.sh       # Downloads both models on first boot, starts Flask
├── api.py              # Flask REST API — the full PDF pipeline
└── README.md

---

Models downloaded automatically on first boot

Model	Size	Source
UTRNet-Large (best_norm_ED.pth)	~300 MB	Google Drive
YOLOv8 Urdu detector (yolov8m_UrduDoc.pt)	~50 MB	GitHub Releases

Both are saved to the persistent volume /app/models — downloaded once, reused forever.

---

Deploying on Coolify

1. Push these 4 files to a new GitHub repo (e.g. utrnet-pdf)
2. Coolify → New Resource → Application → Git Repository
3. Build pack: Dockerfile | Port: 5000
4. Add Persistent Volume:
   • Name: utrnet_models
   • Mount path: /app/models
5. Deploy — first boot downloads ~350 MB (watch logs, takes 2–5 min on typical VPS)
6. Health check: GET /health → {"status":"ok","models":["UTRNet-Large","YOLOv8-UrduDoc"]}

---

API Usage

POST /ocr-pdf

Send a PDF, receive a ZIP of .txt files.

curl -X POST http://YOUR_VPS:5000/ocr-pdf \
  -F "pdf=@/path/to/urdu_book.pdf" \
  -o output.zip

Then unzip:

unzip output.zip
# urdu_book_page_001.txt
# urdu_book_page_002.txt
# ...

Python example:

import requests

with open("urdu_book.pdf", "rb") as f:
    resp = requests.post(
        "http://YOUR_VPS:5000/ocr-pdf",
        files={"pdf": ("urdu_book.pdf", f, "application/pdf")}
    )

with open("output.zip", "wb") as out:
    out.write(resp.content)

GET /health

curl http://YOUR_VPS:5000/health
# {"status": "ok", "models": ["UTRNet-Large", "YOLOv8-UrduDoc"]}

---

Performance expectations (CPU VPS)

Pages	Approximate time
1 page	~15–30 seconds
10 pages	~3–5 minutes
50 pages	~15–25 minutes

CPU inference is slow for large books. For a 200-page Urdu book, consider submitting it as a background job rather than a synchronous HTTP request (or add a queue with Redis + RQ on top of this API).

---

Output format

Each .txt file contains the recognized Urdu text for that page, one line per detected text line, top-to-bottom reading order. Files are UTF-8 with BOM (so Windows Notepad renders Urdu correctly).

---

Notes

• Max PDF upload size: 100 MB
• Pages are rendered at 200 DPI — higher DPI improves accuracy on small text but increases processing time
• The threaded=False Flask setting ensures only one PDF is processed at a time (avoids memory issues on low-RAM VPS)
• License: UTRNet and YOLOv8-UrduDoc are CC BY-NC-SA 4.0 — non-commercial/research use only