OCR
Make scanned PDFs searchable
The OCR engine needs more memory than iOS Safari and in-app browsers (WeChat, QQ, Douyin, Xiaohongshu, etc.) allow per tab, so it can't run here. Please open this page on a desktop browser (Chrome, Edge, Firefox, or Safari on macOS).
How to use
Purpose
Run optical character recognition on scanned PDFs to turn image-based pages into a real text layer — making the PDF fully searchable, copyable, and downstream-extractable. The OCR model runs locally in your browser (first run downloads ~5-35 MB of language data, then works fully offline). No upload, no watermark, free, no page limit. Use it to enable full-text search across scanned contracts, digitize library/archive materials, or convert handwritten note scans into editable text.
Steps
- Drag or pick a scanned PDF
- Pick language: Chinese / English / Chinese+English / Japanese / Korean
- First use auto-downloads the language model (English ~5MB, Chinese ~30MB)
- Wait for the model to load (5-15 seconds)
- (Optional) Toggle "Keep original image as background" (recommended for searchable PDF output)
- (Optional) Tweak "DPI resample" (300 default, balance of accuracy and speed)
- Click "Start OCR"
- Watch the progress bar — roughly 3-8 seconds per page on a modern CPU
- Output is a searchable PDF (text layer over the scan) or plain TXT
FAQ
- Will my files be uploaded?
- No. The OCR model (Tesseract engine + training data) runs locally via WebAssembly; scanned files never leave your device. This is materially different from cloud OCR (Google Vision, Azure OCR) which sends content to a server.
- How accurate is the recognition?
- Print English typically scores 95%+, clean scans approach 99%. Handwriting, blurry scans, complex layouts (tables, multi-column) drop accuracy noticeably. Scan at 300 DPI+ for best results.
- Why is the first run slow?
- First use downloads the language model (English ~5MB, Chinese ~30MB, mixed ~35MB). After that, the model is cached in your browser and ready instantly. Download time varies with your network — typically 3-10 seconds.
- How fast is the recognition?
- Pure-text single A4 page: ~3-5 seconds on a modest laptop (i5/8GB). Mixed image-text or high-res scans: ~8-12 sec/page. For 100-page docs, set it running and grab coffee — typically 10-20 minutes.
- Does it handle handwriting?
- Tesseract has limited handwriting support — accuracy is significantly lower than printed text. Neat block printing might score 70-80%, cursive is essentially unrecognizable.
Use cases
- Make scanned paper contracts searchable for clause-level retrieval
- Digitize library/archive materials for academic full-text search
- Turn phone-shot meeting notes into editable text
- OCR scanned financial reports for downstream data extraction
- Process foreign-language book scans before machine translation
Tip
Scan quality is the single biggest factor: 1) Scan at 300 DPI minimum; 2) Keep pages straight (no skew); 3) Avoid shadows and glare; 4) For black-and-white documents, binarize first for better contrast. Use the Chinese+English mixed model for documents that mix languages — single-language models perform poorly on the other script.