Scrolling Capture OCR on Windows 11: Local Processing, No Cloud

Windows 11’s Snipping Tool does scrolling capture. ShareX does it better. But neither does OCR — and if you use a cloud OCR API, your full-page screenshots leave your machine.

CaptureOCR combines scrolling capture + local OCR in one Windows app. No cloud, no subscription, works offline.

The Scrolling + OCR Gap

Tool	Scrolling Capture	OCR	Local?
Windows Snipping Tool	✅	✅ (Text Actions)	✅	❌ Single region only
ShareX	✅ Excellent	❌	✅	No OCR
Snagit	✅ Excellent	✅ Built-in	❌	Cloud OCR option
PowerToys Text Extractor	❌	✅	✅	Single region only
CaptureOCR	✅ Best-effort	✅ PP-OCRv6	✅	Yes, fully local

How CaptureOCR Does Scrolling Capture

Row-signature stitching (not template matching)

Template matching fails on scrolling: consecutive frames overlap, neither contains the other. cv2.matchTemplate expects template ⊂ target — that’s not scrolling.

Row signatures work better:

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 (the real headache)

Fixed nav bars break naive stitching — overlap region never matches. The matcher retries skipping 8%, 16%, 25% of top rows. If no sticky element, first pass works.

Adaptive wheel stepping

One scroll notch = wildly different pixels across apps. Stepper starts at 1 notch, measures actual shift, targets ~45% selection height, 2× growth cap, 1–12 notch range.

Hard caps: 999 frames, 512 MB canvas, 20,000 px minimum height.

The OCR Side: Tiled PP-OCRv6

Here’s the key insight: a 20,000 px tall image fed to an OCR engine with 1600 px cap gets globally downscaled. 30 px text → 2 px. Noise.

CaptureOCR tiles tall images instead:

• Tile height target: 1600 px
• Overlap: 64 px (2 lines at 1080p)
• Core bands: [0, 1536), [1536, 3072), …
• Each tile extends 32 px 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 exactly once. Overlap ensures boundary lines complete.

Tiles process serially (single ONNX session). Tiling = precision, not throughput.

Real Test Results on Windows 11

Selection	Frames	Height	Stop Reason	Rows Extracted
500×400	30	7,226 px	frame_limit	6 → 183
500×220	60	6,420 px	bottom (natural stop)	10 → 191

The small region (220px tall) scrolled all the way to document bottom naturally. Row numbers increment cleanly — no duplicates, no gaps.

When It Works (and When It Doesn’t)

Works well:

• Terminal output, log viewers (cmd, PowerShell, Windows Terminal)
• Simple web pages with stable scrolling
• Notepad++, VS Code output panels
• PDF viewers with smooth scroll (Edge, Adobe Reader)
• Long chat logs in simple clients

Struggles with:

• Virtualized lists (Discord, Slack, Teams, Twitter/X) — content not rendered until scrolled into view
• Lazy-loaded images — content appears mid-scroll
• Custom scroll animations — frame alignment breaks
• Horizontal scrolling — vertical only
• Complex dynamic pages (SPA with infinite scroll)

Region capture is the reliable fallback. If scrolling fails, capture regions manually.

The Workflow on Windows 11

1. Ctrl+Shift+X → hold Shift → drag scrollable area
2. Click “Start Scroll Capture” in floating panel
3. Watch progress (frame count, canvas height)
4. Stop at bottom or click “Stop”
5. OCR runs automatically (tiled, 1600px tiles, 64px overlap)
6. Export Markdown/CSV/TXT/.snipx

Local Processing = Privacy & Compliance

• No telemetry, no account, no cloud calls
• Tiny model (6.6 MB) bundled in installer — works offline immediately
• Small model (30 MB) downloads once on first use, then runs offline
• Payment via Microsoft Store; we never see payment details
• Settings and sessions in %LOCALAPPDATA%\CaptureOCR\

Performance on Windows 11

Metric	Fast Mode (tiny)	Accurate Mode (small)
Model size	6.61 MB	29.79 MB
Load time	1.35s	1.06s
Per image	~78 ms	~258 ms
P95 latency	144 ms	471 ms
Languages	49 (no Japanese)	50 (includes Japanese)

ONNX Runtime CPU — no GPU needed, no CUDA, no DirectML. Runs on any Windows 10/11 x64.

Compare: Cloud vs Local for Scrolling OCR

Factor	Cloud API (AWS/Azure/Google)	CaptureOCR (Local)
Data leaves machine	Yes (full page screenshots)	Never
Per-page cost	$0.001–0.015	$0 (one-time $2.99)
Internet required	Yes	No (after install)
Setup	API keys, SDK, billing	Install → run
Output	JSON (custom parsing)	CSV/Markdown ready for Excel
Vertical text	Limited	✅ Native
Offline capable	❌	✅

Windows 11 Specifics

• Snipping Tool does scrolling capture but single-region OCR only
• CaptureOCR adds: multi-region batch, structured export, coordinates, confidence
• Works on Windows 10 2004+ (19041) and Windows 11
• MSIX packaging — clean install/uninstall, no registry pollution
• Portable zip available — unzip to USB, run anywhere

The Workflow

1. Ctrl+Shift+X → hold Shift → drag scrollable area
2. Click “Start Scroll Capture” in floating panel
3. Watch progress (frame count, canvas height)
4. Stop at bottom or click “Stop”
5. OCR runs automatically (tiled, 1600px tiles, 64px overlap)
6. Export Markdown/CSV/TXT/.snipx

When Scrolling Fails: Region Capture Fallback

If scrolling capture fails (virtualized list, lazy loading, custom animation):

1. Ctrl+Shift+X (no Shift)
2. Drag regions manually around visible content
3. Scroll → capture next batch → repeat
4. Batch OCR → single export

Slower but 100% reliable.

Get Local Scrolling OCR on Windows 11

1. Install from Microsoft Store ($2.99 one-time, no subscription)
2. Open long page (terminal, log, doc, chat)
3. Ctrl+Shift+X + Shift → capture scroll
4. Export CSV → done

No cloud, no subscription, works offline.

---

---

Related reads:

• How to Extract Text from Multiple Screenshots at Once
• Batch OCR Multiple Regions Single Session Windows
• PP-OCRv6 vs Windows Snipping Tool OCR Comparison