PrivatePDF

Make a scanned PDF searchable

A scanned document is a photograph of words. You can read it, and your computer cannot — so you cannot search it, cannot copy from it, and a screen reader gets nothing. Recognition fixes that, and the documents people most want to scan are the ones they least want to upload.

or drop a file anywhere on this panel

opened locally · 0 bytes sent

How it works

  1. 01Open the scanned PDF.
  2. 02Recognition starts on its own, page by page, and the panel counts as it goes.
  3. 03The pages do not change — the words are added invisibly on top.
  4. 04Export a PDF you can search.

The page still looks exactly the same

Many OCR tools rebuild the document: they re-render every page and hand you a file that looks slightly different, usually slightly worse. This adds the words as invisible text over the scan you already had. Select a line and the selection lands on the ink; look at the page and it is untouched, pixel for pixel. There is a test that renders the page before and after and fails if a single pixel moved.

The engine is served from this site

The recognition engine and its language data are about nine megabytes, downloaded once the first time you use this tool — from here, not from a content delivery network. That distinction is the whole point: the usual setup fetches them from a third party, which would be a request announcing that someone is recognising a document. There is a test that fails the build if any request leaves this origin.

It skips pages that already have text

Running recognition over a page that already contains real text does not improve it — it adds a second, worse copy underneath the first, so every search then returns everything twice. Pages are checked before anything is recognised, and a document that is already searchable is left alone with the panel saying so.

Low-confidence words are dropped

Tesseract will confidently read speckles, page borders and scanner artefacts as words. An invisible layer full of that is worse than no layer at all, because it makes the document searchable for words it does not contain — and a search that returns a false hit is harder to recover from than one that returns nothing. Anything read with low confidence is discarded.

Recognition is never perfect

A clean 300-DPI scan of printed text reads very well. A crooked, faint or speckled one reads poorly, and the panel tells you when confidence is low rather than handing you a searchable document you should not trust. Do not use a search of an OCRed scan to prove something is absent — for that, read it.

Questions

Which languages are supported?
English at the moment. Each additional language is another three megabytes to serve from this site, so they are being added deliberately rather than all at once.
Why is the first run slow?
It downloads the engine and the language data — about nine megabytes, once. After that your browser has them cached and recognition starts immediately.
Can I get the text out as a plain file?
Yes. Once a scan has been made searchable, the Extract text tool will read it like any other document.
Does it work on handwriting?
Barely, and you should not rely on it. The engine is trained on printed text.
Can I OCR a crooked scan?
You can, and it will read worse than it needs to. Straighten it first with Clean up a scan — recognition reads a straightened, de-speckled page substantially better, and it reads speckles as punctuation.