OCR and text recognition, explained
The term OCR appears everywhere scanning is discussed, and it is rarely explained. Here is what sits behind it, where the limits are, and why this one step decides whether you will find your documents again later.
Last reviewed: August 12, 2026
What is OCR?
OCR stands for optical character recognition. It is a process that turns a picture of text back into actual text: the software identifies the individual letters on a scan or a photo and stores them as characters you can search, copy and work with.
You meet OCR constantly, usually without anyone using the word: in the copier that produces a searchable PDF, in the photo app that lets you copy a phone number out of a picture, in the banking app that reads a transfer form. For your paperwork, OCR is precisely the step between "I have a picture of the letter" and "I can find the letter again".
Why a photo alone is no use
When you photograph a letter you create an image: perfectly readable to you, and to the computer a surface of light and dark dots in which it recognises neither a date nor an amount. Finding that photo again later means knowing its file name.
Text recognition closes that gap. It identifies the letters in the image and stores the text alongside, usually invisibly behind the picture. The document looks exactly as it did before. It's searchable all the same.
Forty years of Tesseract
The best-known text recognition in the world is older than the web. Between 1985 and 1994, Hewlett-Packard built a program called Tesseract in Bristol and in Greeley, Colorado. It came out among the top performers at the University of Nevada's OCR contest in 1995, and then it went into a drawer.
In 2005 HP released the source code together with the university, under the Apache licence. Google funded the work from 2006 for roughly a decade; a community of volunteers has maintained it since. Version 4 gave Tesseract a neural network in 2018, version 5 followed at the end of 2021, and both still sit inside countless scanners, archive tools and library systems.
On cleanly scanned book pages with a simple layout, Tesseract still does well, free of charge and with no internet connection. Its weaknesses are just as well known: photos with shadows, nested layouts, tables and handwriting. That's exactly what the second generation went after.
What text recognition does well today
With printed text on ordinary paper, text recognition is now very reliable: invoices, contracts, official letters and bank statements come out almost flawless given a decent capture, including multiple columns and several languages. Errors are the exception.
Modern methods cope well with tables and forms too, and they also recognise where something sits on the page. That's what makes it possible to tell whether a number is the invoice total or a customer reference.
Where it gets difficult
Handwriting is the hardest test. Clear block capitals often work; a hurried note in the margin doesn't always. Expect hits here, not certainty.
Also reliably difficult: very faint thermal receipts, heavily creased paper, captures with a hard shadow or flash reflection, stamped text across the writing, and faxed copies of copies. A better capture solves almost all of it. Better software doesn't.
- Faint till receipts that are already fading
- Shadows, flash and tilted captures
- Creased or curled paper
- Joined-up handwriting
What changed in 2025
Since 2025, text recognition mostly comes from models that learned images and language together. The difference shows day to day: they read more than characters, taking in structure and context as well, so tables, footnotes, multi-column letters and the order you're meant to read them in.
A few names you may run into, with their release dates: Mistral OCR from France (March 2025, a paid API), DeepSeek-OCR from China (October 2025, open weights under an MIT licence), olmOCR from the non-profit Allen Institute for AI (second version October 2025, Apache licence) and dots.ocr, out of the Xiaohongshu orbit. Little of that matters to you directly. It does explain why a photographed till receipt works today where three years ago it would have come out as gibberish.
One word of caution about leaderboards: most accuracy figures you'll read about these models come from the makers themselves, measured on test runs they ran themselves. That isn't disreputable, but it isn't independent verification either. Even the most-cited yardstick, OmniDocBench from OpenDataLab, is usually run by the model providers themselves, not by a neutral party.
The difference between reading and classifying
Plain text recognition gives you a sequence of characters, no more. Whether this is an invoice or a cancellation, whether the date is the invoice date or the due date, it doesn't say.
Language models handle that second step today. They classify the recognised text, suggest a name and identify the type of document. Day to day the difference is very noticeable: with plain text recognition you type the name yourself, with classification you confirm a suggestion.
How to spot text recognition worth having
Check three practical points instead of the marketing claims. First: is the recognised text stored with the document so you can search inside it later? Second: does the search find words in the middle of the text? Third: what happens when recognition gets something wrong, can you correct it?
And one question that weighs just as heavily for personal papers: where does the reading happen? If recognition runs at a provider abroad, your document leaves the country. With health and financial records that deserves a deliberate decision.
Compare it yourself, no expertise needed
If you want to know how good text recognition really is now, try it on a document of your own. On OCR Arena you upload a page, two models process it, and you see both results side by side without being told which came from which. Then you vote, and a public ranking builds up out of those votes.
Part of judging it is knowing who runs it: the site belongs to Extend, a company in the document-processing business, and the vote is a popularity contest instead of a scientific measurement. As a quick practical test on your own paper it still works, and it beats any marketing claim. Take something difficult: a crumpled receipt, a form, a handwritten note.
Frequently asked questions
Is OCR the same as scanning?
No, they are two steps. Scanning produces an image of the page, nothing more. Text recognition comes afterwards and turns that image into readable, searchable text. Many devices and apps do both in one go, which is why the difference tends to go unnoticed.
Does text recognition work on handwriting?
Partly. Clean printed handwriting is often read well, joined-up writing much less so. With handwritten notes, don't rely on every word being findable.
Do I need OCR if I name my documents anyway?
A good file name helps with sorting. But you'll often search for something inside the text: a contract number, an amount, a provider's name. Only text recognition finds that.
Is a PDF automatically searchable?
No. A PDF from a scanner or a camera often contains nothing but an image. It becomes searchable only once text recognition has run over it and stored the text.
Which text recognition is the best?
For printed letters and invoices there's little between the current models. Differences turn up with tables, forms and handwriting, and there the order changes with every new release. What matters more to you than the name of the model is whether the recognised text gets stored permanently and stays searchable.
Can I use Tesseract myself?
Yes. Tesseract is free and runs on your own machine with no internet connection. It's only comfortable if you get along with the command line or a front-end program; on a clean book scan it does a good job, and with phone photos and tables it reaches its limits.