OCR

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

1. Drag or pick a scanned PDF
2. Pick language: Chinese / English / Chinese+English / Japanese / Korean
3. First use auto-downloads the language model (English ~5MB, Chinese ~30MB)
4. Wait for the model to load (5-15 seconds)
5. (Optional) Toggle "Keep original image as background" (recommended for searchable PDF output)
6. (Optional) Tweak "DPI resample" (300 default, balance of accuracy and speed)
7. Click "Start OCR"
8. Watch the progress bar — roughly 3-8 seconds per page on a modern CPU
9. 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.