OCR guide

OCR (optical character recognition) reads the text out of a scan and turns a picture of a page into characters a computer can find. Irfan PDF offers two ways to do it, and they make a different trade: one keeps the page exactly as it looks, the other gives you text you can correct.

OCR can introduce recognition errors

No OCR system is perfect, and this one is not claimed to be. Characters get confused — 0 with O, 1 with l, rn with m — especially on faint, skewed or low-resolution scans. Never rely on OCR alone for legal, medical or financial numbers. Check any number, name or date that matters against the original page before you use it.

How OCR runs here

The engine is RapidOCR (rapidocr-onnxruntime 1.4.4) on ONNX Runtime with the official English PP-OCRv5 recognition model. It runs fully offline on your machine — no Tesseract install, no cloud OCR, no API key.
On upload, scans are auto-oriented (sideways or upside-down pages are turned upright) and automatically deskewed when the measured tilt is between 0.30° and 5.00°, with guards that only apply it to flat scans.
Multi-page scanned PDFs are processed page by page, with per-page progress.

Searchable OCR

Adds an invisible text layer over the page image. The words can be selected, copied and searched, but nothing you see changes — the page still looks exactly like the scan.

Original appearance preserved: word-level placement, size and sampled colour.
Preserved through the tests: signatures, stamps, digits and overall visual appearance, plus page dimensions and layout (test_document_integrity, 52 checks).
The document becomes findable in the library the moment it has a text layer — searching a word inside the scan returns the file with the matching sentence.
Best for: archives, records and anything you want to search or keep looking exactly as it does today.

Editable OCR

Covers the printed glyphs and inserts real, selectable, editable text at their original positions, so you can click a word in the editor and fix it.

Table cells are written back in reading order (row by row, left to right).
Layout changes may occur and formatting may differ. This mode rebuilds text rather than merely tagging it: spacing, line breaks, columns and typeface can come out differently from the original scan, and recognition mistakes are now yours to correct.
Best for: when you must fix words — a wrong name or amount in a form — and you are willing to re-check the layout afterwards.

Mode	Searchable OCR	Editable OCR
What it changes	Nothing you can see — an invisible text layer is added.	The printed glyphs are covered and replaced with real text.
Layout risk	None: appearance is preserved.	Layout and formatting may differ from the original.
Use it when	You want to search and keep the page as-is.	You need to correct the words themselves.

Shared limits — read these before you start

English only. Only the official English PP-OCRv5 model is bundled and verified. An “OCR language” setting exists, but recognition in any other language is Not Verified and must not be relied on.
Errors happen. See the warning above: check anything critical against the original.
Digitally signed PDFs are refused. Adding a text layer would invalidate the signature, so the app explains and refuses instead of quietly breaking it — nothing is written when a file is refused.
Encrypted (password-protected) PDFs are refused with a message that names the fix. Removing a password you know is available as a tool; opening a file you do not have the password for is not implemented.

What the tests check about preservation

Page dimensions and layout are verified by test_document_integrity (52 checks); signatures, stamps, digits and visual appearance through OCR by _test_ocr_fidelity (19), recert_freshness (12), recert2_analysis (11) and _test_font_preserve (24) - all in the 43-suite gate, all 43 passing on the 2026-10-17 run. What is never claimed: 100% signature validity after a modification. That is exactly why signed files are refused instead of edited.

How to check the result

Search two or three words you know are on the page — if search finds them, the text layer is real.
Spot-check names, dates and amounts against the original image, character by character.
For editable OCR, scroll the whole document once: a rebuild can shift columns, line breaks or spacing even when every word is right.
OCR writes a new file — for example invoice (searchable).pdf or invoice (editable).pdf — and leaves the original untouched, so you always have the source to compare against.

Next: Conversion guide (Word, Excel, images) · FAQ · Security