Polish OCR Comparison

Side-by-side OCR comparison for Polish text — open-source engines + major cloud APIs.

Engines

Engine	Type	Polish support
Tesseract	local	pol traineddata
EasyOCR	local (GPU)	pl language pack
docTR	local (GPU)	Latin-script multilingual (Polish diacritics)
Surya	local (GPU)	90+ languages, auto-detect
RysOCR	local VLM (GPU)	Polish LoRA on PaddleOCR-VL
Qwen2.5-VL	local VLM (GPU)	general multimodal LLM as OCR; QWEN_VL_MODEL (default 3B)
Google Cloud Vision	API	language_hints: [pl]
AWS Textract	API	Latin script / auto
Azure Document Intelligence	API	locale: pl-PL
Mistral OCR	API	native Polish support

Local engines work out of the box. Cloud APIs need credentials in .env (see .env.example); they'll show as offline until configured.

Local LLM OCR: rysocr (PaddleOCR-VL + Polish LoRA) and qwen_vl (Qwen2.5-VL instruct). Both are vision-language models prompted for transcription — not classic detect→recognize OCR stacks.

By default OCR_LOCAL_ONLY=1 — only local engines load on first run (faster startup, no API panels). Set OCR_LOCAL_ONLY=0 in .env when you're ready to compare cloud APIs.

Run

cd ~/ocr-webapp
cp .env.example .env   # optional — fill in API keys, set OCR_LOCAL_ONLY=0 for cloud
docker compose up --build

Open http://localhost:8082 (override with OCR_PORT in .env)

First boot downloads model weights (EasyOCR, docTR, Surya, RysOCR, Qwen2.5-VL — cached in Docker volumes).

Usage

1. Drop, paste, or browse an image with Polish text
2. Click Recognize
3. Compare results + latency per engine

Google credentials in Docker

If using Google Vision, mount your service-account JSON:

# docker-compose.yml (add under ocr service)
volumes:
  - ./secrets/gcp-sa.json:/secrets/gcp.json:ro
environment:
  - GOOGLE_APPLICATION_CREDENTIALS=/secrets/gcp.json

Tests

cd ~/ocr-webapp
pip install -r requirements-dev.txt
pytest -v

Integration tests (real engines/APIs) are skipped unless:

RUN_OCR_INTEGRATION=1 pytest -v -m integration

• Default host port 8082 (OCR_PORT in .env) — ASR app uses 8080
• docTR has no dedicated pl checkpoint; it uses a Latin-script multilingual CRNN (works well for Polish text)
• RysOCR downloads PaddleOCR-VL + LoRA adapter on first boot (~several GB); requires transformers==4.57.6
• Max upload size: 20 MB