OCR PDF

A scanner produces a photograph. The resulting PDF looks like a document and behaves like a picture: you can see the words, and your computer cannot, because the file contains a grid of pixels rather than characters. Nothing can search it, no text can be copied from it, no screen reader can read it aloud, and no converter can turn it into an editable file.

That gap matters most in exactly the archives where it is most common. A filing cabinet scanned to PDF is not searchable, which means finding one invoice among four thousand is a manual job. Contracts scanned for the record cannot be checked for a clause. Research papers cannot be quoted from without retyping.

Optical character recognition closes it. The page image is analysed, the shapes are recognised as characters, and the resulting text is written into the PDF as an invisible layer positioned exactly over the words it came from. The page looks identical and has become a document.

Resolution first. Around 300 DPI is the sweet spot for text — enough detail to distinguish similar letterforms without producing needlessly large files. Below about 200 DPI accuracy falls away quickly, because the fine strokes that separate one character from another stop being captured at all.

Then contrast and geometry. Clean black text on white recognises far better than grey text on a tinted background or a photograph taken at an angle. Skew is handled automatically, as is page rotation, and both help materially — recognition works line by line, so a page tilted a few degrees is much harder than the same page straightened.

The content itself sets the ceiling. Ordinary prose in a standard typeface is the easy case. Dense tables, multiple columns, decorative fonts, mathematical notation and text over images are all harder. Handwriting is a genuinely different problem and standard OCR does not solve it, however good the scan.

Recognition uses far more than letter shapes. Each language model brings a character set, expected letter combinations and word patterns, all of which resolve the ambiguities that shape alone cannot — the difference between rn and m, between 1 and l, between O and 0 depends on what would make sense in the surrounding word.

Choosing the wrong language therefore degrades every line, not just the accented characters. A German document processed as English loses the umlauts and misreads the long compounds; a Russian or Chinese document processed as English produces nothing usable at all, since the characters are not in the model.

For genuinely bilingual documents, pick the language that carries most of the body text. Names, addresses and short quotations in another language usually survive; a document that is half and half will lose accuracy whichever way you choose.

Order matters more than it might seem. OCR before compressing, because compression resamples the page images and softer images recognise less reliably — and the text layer itself adds almost nothing to file size, so nothing is lost by doing it in that sequence.

OCR before converting, too. PDF to Word, PDF to Excel and PDF to Markdown all need text to work with; given only images they can only hand back images. Adding the text layer first is what makes an editable result possible at all.

And OCR after redacting rather than before, on any document that needs both. Redaction flattens the pages it touches, which destroys any text layer on them — running OCR afterwards restores searchability while reading only what is now visible, so the redacted content stays gone.