Get the words out of a picture

What OCR really is, in plain terms

A photograph of a page contains no words. It contains pixels arranged in shapes that your eye assembles into letters. The computer sees none of that. To it, a scan of a contract and a photograph of a beach are the same thing: a grid of coloured dots.

Optical character recognition is the process of turning those shapes back into characters. The engine finds the lines, splits them into probable letters, compares each shape against what it has learned that language looks like, and produces its best guess along with a measure of how sure it feels. That last part matters, and it is why the result here arrives with a confidence score rather than being presented as fact.

Once the words exist as text, everything becomes possible again. You can search them, edit them, translate them, paste them into an email or feed them to a screen reader. Trapped inside a picture, they can do none of that.

When people reach for this

• A screenshot with information you need. An error message, a receipt, a table of numbers, a message thread. Retyping it is the only alternative and nobody enjoys that.
• A photographed page. A textbook paragraph, a recipe from a cookbook, a page from a manual, a notice on a wall.
• Scanned paperwork. Invoices, contracts, forms and letters that arrived as pictures and now need to be searchable.
• Text inside a graphic. An infographic, a slide, a poster, a screenshot of a chart where the labels are what you actually want.
• Study and research. Pulling quotes out of a photographed journal article or a library book you cannot take home.
• Accessibility. Text inside an image is invisible to a screen reader until somebody extracts it.

What decides whether it works

Accuracy is settled almost entirely by the picture you feed in, not by the engine. Understanding what helps and what hurts saves a lot of frustration.

The single biggest factor is size

Recognition needs enough pixels to tell a lowercase L from a one and a zero from a capital O. As a rough guide, letters want to be at least twenty pixels tall. A screenshot taken at normal zoom often falls short of that, which is why this tool enlarges small pictures and stretches the contrast before reading begins. If you are capturing a screenshot yourself, zoom the page in first. Larger letters at capture time beat any amount of processing afterwards.

Getting a clean read

A few habits turn mediocre results into reliable ones.

• Crop tightly. Cut away the desk, the margins, the page number and anything decorative. The engine wastes effort on non text and sometimes invents characters from patterns.
• Straighten the page. Lines should run horizontally. Even a five degree tilt costs accuracy, and most phone gallery apps have a rotate control that fixes it in seconds.
• Shoot parallel to the paper. Standing over the page rather than leaning across it keeps the letters the shape they were printed in.
• Use even light and no flash. Daylight near a window beats a ceiling bulb directly overhead, which throws your own shadow onto the page.
• One column at a time. A two column page gets read across the lines in the order they are found, which scrambles the meaning. Crop each column separately.
• Choose the right language. Picking the wrong one does not produce an error. It produces confident nonsense, because the engine matches shapes against the wrong alphabet.

Reading the confidence score

Every result comes with a percentage. It is the engine reporting how certain it felt about the characters it chose, and it is worth paying attention to.

Above eighty, the text is generally dependable and any errors will be isolated characters. Between sixty and eighty, read it through against the picture before you use it anywhere important. Below sixty, something about the source is fighting the engine: the letters are probably too small, too faint or too crooked. Rather than correcting dozens of mistakes by hand, go back and get a better picture. Two minutes of recapturing beats twenty minutes of proofreading.

Common confusions to look for even in high scoring results are the digit one against a lowercase L and a capital I, zero against a capital O, five against S, and rn appearing where the original had an m. If the text contains reference numbers, prices or codes, those are exactly the places to check first.

Languages and the first slow run

Each language has its own trained data, and the file for the language you choose downloads the first time you use it. That means your first run in a new language takes noticeably longer than every one after it, because the browser is fetching the data before it can start reading. Once it is cached, subsequent jobs in that language begin immediately.

Languages that share the Latin alphabet, such as English, Spanish, French, German, Portuguese and Italian, tend to be the most accurate, because the shapes are well covered and the training data is large. Arabic and Urdu read right to left and join their letters, which is inherently harder. Chinese, Japanese and Korean have far larger character sets, so their data files are bigger and recognition asks more of your device. All of them work, but expect a difference in speed and in how much proofreading you do afterwards.

Where this cannot help

Being clear about the limits is more useful than pretending they do not exist.

• Handwriting. The engine learned printed characters. Neat block capitals occasionally survive, but ordinary joined up writing will not, and no setting changes that.
• Layout. Reading order is preserved and line breaks can be kept, but columns, tables, sidebars and text boxes are flattened into ordinary lines. A complicated page needs cropping into pieces.
• PDF files directly. A PDF page is not an image until something draws it. If yours is a scan, screenshot the page and paste that here. If it already contains real text, you do not need OCR at all: select and copy it in any reader.
• Text hidden by damage. Nothing can recover characters that a fold, a stain or a heavy crease removed from the page.
• Perfect output, ever. Every OCR result deserves a read through. Treat it as a very fast first draft rather than a finished transcription.

Why doing it in your own browser matters here

Think about what people actually put through an OCR tool. Payslips. Bank letters. Passport pages. Medical results. Signed contracts. Identity documents for a visa application. These are among the most sensitive files an ordinary person owns, and the typical free OCR site works by uploading them to a server, processing them there and returning the text.

That means your document sat on hardware you cannot inspect, for a length of time you have to take on trust, handled by a company whose retention policy you have almost certainly not read.

This page never sends the picture anywhere. Your browser downloads the recognition engine and the language file, and after that all the work happens inside the tab on your own machine. Open the network panel in your developer tools and watch while it runs if you want to see it for yourself. A practical side effect is that once the page and the language have loaded, the tool keeps working even if your connection drops.

What to do with the text once you have it

The result lands in a box you can edit directly, so fix the two characters it misread rather than retyping the paragraph. Word and character counts update as you work, which is handy when you are pulling a quote to a length limit. From there you can copy it straight into whatever you are writing, or save it as a plain text file that opens in anything.

If the picture you started with is also headed somewhere with a size limit, run it through our image compressor afterwards. Scanned paperwork that needs to travel by email usually wants the PDF compressor instead, and the rest of the text and speech tools handle the jobs that come next.