Make a scanned PDF searchable

An invisible layer over an unchanged page

The scan is kept exactly as it was. What is added is a layer of text drawn in an invisible rendering mode — real characters, positioned word by word over the words in the picture, marked so that they are never painted. The page looks identical because visually nothing changed; select a sentence and the highlight lands on the right words because the invisible text is where the visible ink is. This is the standard way a searchable scan is made, and it is why the result can be searched, copied and indexed while still being a photograph of a document.

Two things about the page are deliberately corrected, because the reading needed them anyway. A page scanned sideways is turned upright, and a page that went through the feeder crooked is straightened — so the output is often tidier than the input. Page sizes are kept from the original PDF rather than normalised, so a document of mixed page sizes stays that way and nothing is rescaled.

Accuracy depends almost entirely on the scan, and the numbers are not close. A clean 300 DPI scan of printed text reads at very high accuracy; the same page at 150 DPI reads noticeably worse; a photograph taken at an angle in poor light is a different proposition entirely. If you control the scanning, 300 DPI in greyscale is the setting that matters — more resolution than that rarely helps, and less hurts immediately. Small print, tight line spacing, coloured backgrounds, photocopies of photocopies and anything stamped or handwritten are all harder in ways no software fully solves.

What is deliberately left out is as important as what goes in. Photographs, logos, signatures and decorative graphics are ignored rather than read, because forcing letters out of a picture produces plausible nonsense that then appears in every search. Chinese, Japanese and Korean words are recognised but left out of the text layer and counted for you, rather than written as blanks, until a font that can carry them ships. Nothing here re-typesets the document: the words on the page are still a picture, and converting to Word is the tool for turning them into editable text.

When you want something else

OCRmyPDF is the tool this page is a convenience next to, and it is free. ocrmypdf --deskew --rotate-pages in.pdf out.pdf does the same job with more control, adds --optimize 3 to shrink the result, --redo-ocr to replace a bad existing layer, and --output-type pdfa to produce an archival file in the same pass. It handles thousands of pages, runs in a folder watcher, and is what libraries and document archives actually use.

For difficult material — historical print, unusual layouts, heavy tables — a hosted service from Google, Amazon or Microsoft will read pages that general-purpose recognition cannot, because they are trained on far more material. That is a genuine capability difference and worth knowing about. It also means sending your documents to a third party, which is exactly the trade this page exists to avoid — so it is the right choice for a fragile old book and the wrong one for a box of medical records.