OCR a scanned PDF

How to make a scanned PDF searchable

1. Choose the language the document is written in.
2. Leave Clean each scanned page first ticked unless the scan is already crisp and white.
3. Choose your PDF, or drop it on the page. It is processed in your browser and never uploaded.
4. Wait while each page is cleaned and read. The progress bar shows the page and a rough time remaining.
5. Try the search box on the results screen, compare the original with the result, and download the new PDF.

What you get

• A PDF that looks like the cleaned scan, with an invisible text layer on top.
• Text you can search, select and copy in any PDF viewer.
• Pages that already contain real text, with no images, left exactly as they were.
• A word count and the engine's own confidence figure, so you can judge how well it read the document.

Why the page is cleaned first

Recognition engines are trained on clean, high-contrast text. A faded or grey scan makes them miss words or misread them. On the faded demo scan used on this site, the recognition engine found 6 of 24 test words in the original and 18 of 24 after cleaning. That is one synthetic sample in English, so treat it as an illustration rather than a promise; but the direction is consistent with how these engines behave.

The cleaning is the same engine as Enhance a scanned PDF: the paper is whitened, shadows are flattened and faint ink is darkened, page by page.

Languages and limits

• Languages: English, French, German and Spanish. Other Latin-script languages are not available yet, and neither are Arabic, Urdu, Hindi, Chinese, Japanese or Korean.
• Pages: up to 20 scanned pages per run. A longer PDF can be split and done in parts.
• Speed: OCR is the slowest thing on this site. Expect several seconds per page, and longer on a phone or an older computer. On a computer with plenty of memory two pages are read at once.
• First use: the engine and the language model are a few megabytes. They are downloaded from this site the first time and then cached by your browser.

What OCR cannot do

• It makes mistakes, especially with numbers, names, small print and unusual fonts. Check anything important against the page.
• It cannot read handwriting.
• It does not rebuild tables, columns or layout: the text layer is the words in position, which is what search and copy need.
• It does not straighten crooked pages. A page scanned at an angle reads less accurately.
• It does not produce an archival PDF/A file.

What happens to the page images

Scanned pages are rendered at 200 DPI and written back as images, with the text layer on top. If your original scan was captured at a higher resolution, the new file is at 200 DPI, so keep the original. The new file is not modified in any other way than cleaning and the added text.

Privacy

The recognition engine, its worker and the language models are files on this site, loaded by your browser. No request carries your document, the page images or the recognised text anywhere. The language model is kept by your browser's normal cache; the text of your document is not stored by the page.