Best Offline OCR for Chinese Invoices on Windows 11
If you process Chinese vendor invoices, you know the drill: screenshot each field, paste into a translator, manually type into Excel. Or worse — upload sensitive financial docs to a cloud OCR API.
There’s a better way. CaptureOCR runs PP-OCRv6 locally on Windows via ONNX Runtime. No cloud, no subscription, no data leaving your machine.
Why Cloud OCR Fails for Chinese Invoices
Most cloud OCR services (AWS Textract, Azure Form Recognizer, Google Document AI) have three problems for Chinese invoice workflows:
- Data sovereignty — financial docs can’t leave your network per company policy
- Mixed script accuracy — Chinese vendor name + English address + numbers + currency symbols = boundary errors
- Cost at scale — $0.001–0.015/page adds up fast for monthly invoice batches
- Internet dependency — no OCR when the network’s down
Windows Snipping Tool’s built-in OCR is free and local, but it’s single-region only and struggles with vertical text, rare characters, and small print.
How CaptureOCR Handles Chinese Invoices
Multi-region capture = one invoice, one session
Open the invoice (PDF in Edge, scanned image, web portal — CaptureOCR captures pixels, not file formats). Press Ctrl+Shift+X and drag regions around each field:
- ① Invoice number
- ② Date
- ③ Vendor name (Chinese)
- ④ Line items table area
- ⑤ Subtotal
- ⑥ Tax
- ⑦ Total
Each region auto-numbered ①②③. Reorder in the drawer if needed.
Two OCR modes for Chinese
| Mode | Model | Speed | Chinese Support | Best For |
|---|---|---|---|---|
| Fast | tiny (~6.6 MB) | ~78 ms/img | Simplified + Traditional + 46 Latin scripts | Clean printed invoices |
| Accurate | small (~30 MB) | ~258 ms/img | + Japanese, better on rare chars | Low contrast, small print, mixed scripts |
Both run 100% locally via ONNX Runtime. Tiny model bundled in installer — works immediately after install. Small model downloads once (~30 MB) then runs offline forever.
Real Chinese invoice test
Input: Typical vendor invoice with:
- Chinese vendor name + English address
- Mixed currency symbols (¥ and $)
- Small-print terms at bottom (8pt)
- 12-row line items table
CaptureOCR output (Fast mode, 5 regions, ~390ms total):
| Region | Field | Confidence | Sample Output |
|---|---|---|---|
| ① | Invoice # | 0.998 | INV-2024-0315 |
| ② | Date | 0.992 | 2024-03-15 |
| ③ | Vendor | 0.987 | 北京华夏科技有限公司 |
| ④ | Line items | 0.981 | Widget A × 100 @ ¥25.00 Widget B × 50 @ ¥45.00 |
| ⑤ | Total | 0.999 | ¥4,750.00 |
The line items region captures the whole table as text lines. Each row becomes a separate CSV row with coordinates. Pivot in Excel for analysis.
Export: CSV That Opens Correctly in Excel
Export → CSV (UTF-8 with BOM) → open in Excel. Every row includes:
- Region index (1, 2, 3…)
- Screen coordinates (x, y, w, h) — physical pixels, DPI-independent
- Line number within region
- Confidence score (0.0000–1.0000)
- Recognized text (UTF-8)
- OCR mode used
- Processing time
- Model hash (for reproducibility)
UTF-8 BOM means Chinese characters display correctly in Excel without the “Data → Text to Columns” dance.
Privacy & Compliance for Finance
- No telemetry — zero usage data collected
- No account required — install and run
- Payment via Microsoft — we never see payment details
- Local storage only — settings, sessions, images in
%LOCALAPPDATA%\CaptureOCR\ - Works fully offline after install (tiny model bundled)
When to Use Accurate Mode for Invoices
Switch to Accurate (small model) when:
- Invoice has low contrast (light gray text on white)
- Text is unusually small (under 10pt)
- Mixed Chinese/English/Japanese on same page
- Dot-matrix or unusual fonts
The 30 MB download happens once. After that, it’s just as offline as Fast mode.
Compare: Cloud API vs CaptureOCR for Invoices
| Factor | Cloud OCR APIs | CaptureOCR |
|---|---|---|
| Data leaves machine | Yes | Never |
| Per-page cost | $0.001–0.015 | $0 (one-time $2.99) |
| Internet required | Yes | No (after install) |
| Setup complexity | API keys, SDK, billing | Install → run |
| Output format | JSON (custom parsing) | CSV/Markdown ready for Excel |
| Chinese/English mixed | Good (extra cost) | Native PP-OCRv6 support |
| Vertical text | Limited | Supported |
| Rare characters (生僻字) | Hit or miss | 7,000+ char coverage |
Get Started
- Install from Microsoft Store ($2.99 one-time, no subscription)
- Open invoice on screen (PDF, browser, scanned image — any app)
Ctrl+Shift+X→ capture 5–7 regions- Export CSV → drop into your accounting workflow
Related reads: