PP-OCRv6 Windows ONNX Runtime Setup: Local OCR Without Python Dependencies

If you’ve tried running PaddleOCR on Windows, you know the pain: Python environment, pip dependencies, version conflicts, model downloads, GPU/CUDA setup. CaptureOCR solves this by bundling PP-OCRv6 with ONNX Runtime in a native Windows app. Zero Python, zero dependencies, runs out of the box.

The Python OCR Problem

Typical PaddleOCR setup on Windows:

1. Install Python 3.10+
2. pip install paddlepaddle paddleocr (huge, pulls CUDA if not careful)
3. Download models manually or let rapidocr handle it
4. Deal with OpenCV DLL conflicts (opencv-python vs opencv-python-headless)
5. ONNX Runtime providers: CPU vs CUDA vs DirectML — which works?
6. Model version drift: rapidocr updates break model compatibility

CaptureOCR sidesteps all of this by:

• Pre-converting PP-OCRv6 to ONNX at build time
• Bundling ONNX Runtime (CPU, no CUDA needed)
• Shipping models inside the installer (tiny = 6.6 MB, no download)
• Managing models via custom ModelManager (SHA256 verified, resumable, ModelScope mirror)

Architecture: ONNX Runtime + PP-OCRv6

CaptureOCR (PySide6/Qt)
    → mss capture (physical pixels, BGRA numpy)
    → Preprocess (resize, normalize, HWC→CHW)
    → ONNX Runtime InferenceSession
        → Det model (text detection)
        → Rec model (text recognition)  
        → Cls model (orientation, optional)
    → Postprocess (box grouping, line ordering, confidence)
    → Structured output (Markdown/CSV/JSON)

Zero Python at runtime. The inference runs in a dedicated worker thread via ONNX Runtime C API (via onnxruntime Python package at build time, but the packaged app is pure native + ONNX Runtime DLLs).

Model Management: Zero-Hassle

Two tiers, both local

Model storage

%LOCALAPPDATA%\CaptureOCR\SnipOCR\models\PP-OCRv6\
├── tiny\
│   ├── det.onnx
│   ├── rec.onnx
│   └── cls.onnx (optional)
└── small\
    ├── det.onnx
    ├── rec.onnx
    └── cls.onnx

Download: SHA256 verified, resumable, ModelScope mirror

# infra/model_manager.py
_MS = "https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/v3.9.2"
_ONNX = f"{_MS}/onnx/PP-OCRv6"

• SHA256 verified on every file
• Resumable download (Range headers)
• Falls back to alternative URLs
• Progress callback for UI

Fast mode = zero download

Tiny model (6.61 MB) is inside the MSIX/installer. Install → run → OCR works. No network call.

Accurate mode downloads small model once (~30 MB, ~1 second on broadband). After that, fully offline.

Performance Tuning: 4 Choices That Mattered More Than Model Choice

1. Never upscale before detection

PaddleOCR default: limit_type="min" + limit_side_len=736 → short side enlarged to 736px.

Problem: 740×92 screenshot → 5920×736 before detection. Huge tensor, slower, no accuracy gain.

Fix: limit_type="max" + det_limit_side_len=1600. Only downscale if long side > 1600px.

• 740×92 stays 740×92 → 8.9× faster (738ms → 85ms)
• CAR 98.59% → 99.86% (simultaneously more accurate)

2. Disable orientation classification

use_cls=False. Classifier trained for 0°/180° rotation; screenshots are upright. Extra stage, no signal.

3. Single inference thread

ONNX Runtime InferenceSession not guaranteed thread-safe. Instead of locks:

• All model loads, unloads, run() calls on single worker thread
• UI enqueues work, receives results via queue
• Cost: no parallel OCR (10 regions = sequential)
• Benefit: predictable switching, no UB

4. Keep fast model hot, unload accurate

• Tiny stays loaded (resident)
• Small unloads after 5s idle → drops resident memory substantially
• CPU memory arena disabled → predictable unload for tray app

Tiling for Long Images: Precision, Not Throughput

A 20,000px tall image fed to OCR with 1600px cap gets globally downscaled. 30px text → 2px. Noise.

CaptureOCR tiles instead:

• Tile height target: 1600px
• Overlap: 64px (2 lines at 1080p)
• Core bands: [0, 1536), [1536, 3072), …
• Each tile extends 32px above/below for context
• Lines assigned by center Y to exactly one core band
• Boxes remapped to whole-image coords, sorted by Y then X

Core bands contiguous → every line claimed once. Overlap ensures boundary lines complete.

Cost: Serial inference 1520ms → 4000ms for 5550px image. Tiling = precision feature, not throughput.

Scroll Capture: Row Signatures, Not Template Matching

Template matching fails: consecutive frames overlap, neither contains the other.

Row signature approach:

1. Grayscale + ignore right 8% (scrollbars)
2. Compress each row → 32 column-bucket means: (H, 32)
3. Shift search: minimize mean absolute difference prev[s:] vs cur[:H-s]
4. Peak validation: best score < threshold AND clearly better than median

Why 32 buckets? Tested on real frame pairs:

• 12 buckets: 7/9 correct
• 24 buckets: 8/9 correct
• 32 buckets: 9/9 correct
• 64 buckets: 9/9 (2× compute)

Sticky headers: Matcher retries skipping 8%/16%/25% of top rows.

Adaptive wheel step: Start at 1 notch → measure actual shift → target 45% selection height, 2× growth cap, 1–12 notches.

Model Verification: SHA256 on Every Build

Build scripts verify model integrity:

# Verify model files match expected SHA256
python scripts/verify_models.py

• Compares every .onnx file against known hashes
• Fails build if mismatch
• Ensures no corrupted/partial downloads in package

Building Your Own: Key Dependencies

If you want to replicate locally:

# Python 3.10+
py -3.12 -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e ".[windows,dev]"

# Critical: use opencv-python-headless, NOT opencv-python
# Otherwise two ffmpeg DLL sets (55 MB extra)
pip install opencv-python-headless
pip install --no-deps rapidocr

ONNX Runtime: onnxruntime (CPU) or onnxruntime-gpu (CUDA) — CaptureOCR uses CPU for portability.

MSIX packaging: Uses makeappx.exe from Windows SDK. --onedir mode (faster startup, models don’t re-extract).

Troubleshooting Common Issues

Why Not Just Use rapidocr Directly?

Get It Running

1. Install from Microsoft Store ($2.99 one-time)
2. Ctrl+Shift+X → capture
3. Done. No Python, no pip, no model downloads (Fast mode).

Related reads:

• PP-OCRv6 vs Windows Snipping Tool OCR Comparison
• Offline OCR Tool That Works on Windows Without Internet
• Best Offline OCR for Chinese Invoices on Windows 11