Make a scanned PDF searchable
Load your PDF, pick German or English and start the recognition. Every page is rendered at 300 dpi, read by OCR and rebuilt with an invisible text layer — the result looks exactly like before, but you can search, select and copy it. Everything runs in your browser, with no upload.
What happens to the pages?
Two clearly separated paths:
- Recognised pages become an image of the page with invisible text on top. The text sits where the letters are in the image, it is simply not painted.
- Skipped pages are copied losslessly — text, fonts and vector graphics stay exactly as they were.
This is the same honest trade-off as compression: whatever goes through rasterisation loses its vector text. That is why the tool checks before you start which pages already have a text layer and offers to leave them alone. For mixed PDFs (a cover page from Word, the rest scanned) that is exactly right.
Does the file get bigger?
Usually yes, often noticeably. A plain black-and-white scan at 200 dpi grows when it is re-encoded as a 300 dpi JPEG, and the text layer adds a little more. The tool shows the size before and after — if the result is too large, send it through the PDF compressor afterwards.
Why 300 dpi?
Because OCR needs pixels. The Tesseract documentation ("Improving the quality of the output") names 300 dpi as the practical lower bound and recommends upscaling smaller type (below roughly 10 pixels x-height) first. Less resolution saves time but costs exactly the accuracy you came for.
Why a limit of 50 pages?
Because every page is rendered and recognised on your own device. Expect several seconds per page — more on a phone. Beyond 50 pages the tool says so openly instead of running the tab out of memory. Split larger documents with Split PDF first and process the parts one after another.
The job runs page by page with a progress indicator and can be cancelled at any time; the recognition worker is shut down cleanly when you do.
Honest limits
- A search aid, not a transcript. Check figures, names and totals before relying on them.
- No handwriting, no blackletter — the models are trained on modern printed type.
- No deskew. Crooked scans stay crooked and are recognised less reliably.
- No layout preservation in the text layer: multi-column pages can come out in the wrong order when copied.
- Encrypted PDFs have to be unlocked first.