SPINAI OCR

The fastest, smallest Korean OCR. ~80 MB models · ~14 ms/line on CPU · fully offline. No GPU, no cloud.

A Korean-first OCR engine built to run on a CPU, offline, in a tiny footprint — for on-prem, air-gapped, and edge deployments where you can't ship data to a cloud API or run a GPU.

🔴 Live demo: https://spinai-ocr-production.up.railway.app — try it in your browser, no install.

---

Why SPINAI OCR?

PaddleOCR and EasyOCR are excellent multilingual giants. Naver CLOVA and PaddleOCR-VL are more accurate on hard Korean text. SPINAI is the opposite trade-off: a small, fast, fully-offline Korean engine for when accuracy-at-any-cost isn't the constraint — data residency, latency, and footprint are.

	SPINAI	EasyOCR	PaddleOCR	Naver CLOVA
Model size	~80 MB	~100 MB+	~100 MB+	cloud
CPU latency / line	~14 ms	~50 ms	varies	network RTT
Offline / no GPU	✅	✅	partial	❌ cloud-only
Docker image	<100 MB	larger	larger	n/a
Korean accuracy	competitive	competitive	higher	highest

Honest about accuracy: if you need maximum accuracy, use PaddleOCR-VL or Naver CLOVA. If you need Korean OCR that runs offline on a CPU in <100 MB, that's us.

Install

pip install git+https://github.com/SpinAIceo/spinai-ocr.git
# PyPI release coming soon: pip install spinaiocr

The Korean model (~80 MB) is downloaded automatically on first run, then everything runs offline. No manual setup — the model is hosted on HuggingFace (spinaiceo/spinai-ocr-consumer-v1).

Quickstart

from spinaiocr import OCRPipeline

ocr = OCRPipeline(lang="ko")          # CPU by default, fully offline
result = ocr("your_image.jpg")
for line in result.lines:
    print(f"{line.confidence:.2f}  {line.text}")

More in examples/. REST API + web UI also included (spinai_ocr.serve).

How it works

• Detection: DBNet (text-region segmentation)
• Recognition: SVTRv2 + CTC (compact, CPU-friendly, no autoregression)
• Languages: Korean (한국어) first-class; English supported

Honest limitations

• Accuracy is competitive, not #1 — PaddleOCR-VL / CLOVA win on decorative fonts, handwriting, and complex scenes.
• Detection can merge regions on cluttered real-world scenes.
• This is the speed/size/offline OCR. We publish where we lose, not just where we win.

License

Apache-2.0. Built with ❤️ for the Korean OCR community. Contributions welcome.