📊
Japanese Contract OCR Benchmark — API vs Open Source, Which Is Better?
Google Vision, Azure, PaddleOCR, Sarashina2.2 — accuracy, speed, price comparison on Japanese contracts
Test Conditions
Input: Japanese real estate contract with mixed vertical/horizontal text, specialized terms, special characters.
Expected Performance
| Tool | CER | Vertical | Speed |
|---|---|---|---|
| Google Vision | ~1-2% | Yes | ~1s |
| Sarashina2.2 | ~1-3% | Yes | ~5s |
| PaddleOCR | ~3-5% | Yes | ~2s |
| EasyOCR | ~5-10% | Partial | ~3s |
Cost (10K pages/month)
| Method | Monthly Cost |
|---|---|
| Google Vision | ~$15 |
| PaddleOCR | $0 (server only) |
| Sarashina2.2 | $0 (needs GPU ~$50/mo) |
Selection
| Situation | Recommendation |
|---|---|
| Confidential docs | Sarashina2.2 (local) |
| Accuracy first | Google Vision |
| No GPU, free | PaddleOCR |
Key Concepts
1
Google Vision — highest accuracy (CER ~1-2%) but data sent to external servers
2
Sarashina2.2 — Google-level accuracy + local execution possible. GPU required
3
PaddleOCR — free without GPU. CER ~3-5% practical. Weak on special characters
4
Confidential docs (contracts) → local execution (Sarashina2.2/PaddleOCR) recommended
Use Cases
Japanese real estate contracts — vertical text + specialized terms + special characters
OCR tool selection — compare tools by accuracy, cost, security for your document type
Security compliance — local for confidential, API for non-confidential, split operation