How to recognize text in a scanned PDF using OCR

Text recognition in a PDF: what happens to a scan

A scanned document looks like a document and reads like a document, but to any program it is a picture. You cannot search it, you cannot copy a contract number out of it, and an intake form that checks for text refuses it as an empty file.

Recognition solves exactly that, and it does so in a way worth understanding: it rewrites nothing.

What actually gets added to the file

The page image is left untouched. A second layer goes over it: text positioned where the corresponding letters sit in the picture, and made invisible.

Every property of the result follows from that. Visually the file is indistinguishable from the original: same paper, same stamps, same tilt if there was one. Meanwhile search finds words, a double click selects a line, and copying yields real text.

The same fact explains why recognition is safe for documents whose appearance matters. A scanned contract with a signature is, after OCR, the same scan with the same signature; it has simply become searchable.

Pages that already hold text are skipped

Before processing, every page is checked for a text layer, and pages that have one are not recognised at all.

Two practical consequences follow.

Mixed documents are handled correctly without any effort. If part of a file was typed in an editor and part was pasted in as a scan, only the second part is recognised. The typed text stays original rather than being replaced by a recognised copy, which matters: original text is exact, recognised text contains errors.

Running OCR on an ordinary text PDF does almost nothing and damages nothing. That is convenient when you do not know whether the document in front of you is a scan: just put it through the tool.

The same logic governs how long a job takes. It depends not on the number of pages in the document but on the number of pages without text. A two-hundred-page report with five pasted-in scans is processed in the time five pages take.

The three quality modes are three resolutions

The quality choice controls one thing: the resolution at which a page is rendered before recognition.

The default is the fast mode, which is worth remembering. A hundred dots per inch is enough for headings and body text at twelve points, but small footnotes, figures in tables and stamps come out noticeably worse at that rendering.

The rule is simple: if the document holds anything you squint at yourself, choose the accurate mode. It is slower, and it saves you doing the job twice.

Language: auto means two languages at once

English and Russian are available, along with an auto mode.

The automatic mode does not guess the language. It recognises the document in both languages at once, which suits text where one language is peppered with names, codes and terms from the other. The price is speed and a small loss of accuracy on uniform text.

If the document is entirely in one language, choosing that language gives a slightly better and faster result.

When the chosen language is unavailable, recognition does not fail: it runs with what is installed and attaches a separate warning naming both the language you asked for and the one applied. That warning is worth reading rather than dismissing: a silently substituted language explains a strange result better than any guesswork.

A note on documents in other languages. Latin script is recognised by the English pack, but diacritics are lost or mangled in the process: French and German texts come out readable with errors in the accents. Writing systems other than Latin and Cyrillic are not supported at all.

How to run recognition

1. Open ocr-pdf and upload the document. 2. Choose the language. Auto for mixed text, a specific language for uniform text. 3. Set the accurate mode if the document holds small type. 4. Run the job and download the result. 5. Open the file and search for a word you know is on the scan. That is the fastest possible check.

Deskewing and what it costs

A separate switch straightens tilted pages before recognition. It is off by default, and there is a reason.

The straightening is applied to a temporary copy of the page, which is what recognition reads, and the text layer is then laid over the original, uncorrected page. On a badly tilted scan that produces an offset: the invisible text ends up rotated relative to the visible letters.

Search does not suffer, the word is still found. Mouse selection does: the selection box lands away from the letters, and copying a passage becomes awkward.

The trade-off is deliberate, because the original pages stay untouched. If the tilt is severe, the better move is to straighten the document before recognition, with rotate-pdf for right-angle turns or by rebuilding the scan through scan-to-pdf, which has real deskewing.

What recognition does not do

It does not improve the source. A blurred photo, a lamp glare across a line, the shadow of a hand over half a page: all of it stays, and recognition over such a spot yields gibberish or skips it.

It does not read handwriting. The tool is built for printed type.

It does not produce legally meaningful text. The recognised layer contains errors and cannot serve as a copy of the document. It exists for search and for systems that require a text layer, not for quoting.

And it does not restore resolution. If the source scan was made at 75 dots per inch, the accurate mode renders it at 300, but no extra letters appear: a blurred picture simply gets bigger. Recognition quality is set by the source, not by the setting.

Where recognition belongs in a chain of operations

The order matters more than it seems, and there is one rule: recognition goes before compression.

Recognition works from the page image. Compression through compress-pdf degrades that image: it lowers the resolution and adds artefacts around the letters. The reverse order produces noticeably more errors in the text, and afterwards there is nothing left to fix them with.

Assembly and rotation belong before recognition too: they change which pages exist and how they are oriented, and the text layer is tied to a specific page.

If what you need is not a file with a text layer but the text itself as a separate file, pdf-to-text exists for that: it extracts the content into a plain text file broken up by page.