How to copy text from a picture

Reading text out of a picture is one of those jobs that is either almost free or not worth starting, and which one it is depends entirely on how the picture was made. This is how to tell the difference before you spend ten minutes on it.

Try to select the text first

Before any recognition happens, try selecting the words on screen. If you can drag across them and they highlight, you already have the text and nothing needs to be read. That covers far more cases than people expect: a screenshot of a web page, a message in almost any app, a digital PDF, an error message in a terminal.

Recognition exists for the cases where selection is impossible — a photograph of a printed page, a picture of a sticker or a whiteboard, a receipt in your hand, a scan of something old. In those cases the words only exist as shapes, and something has to look at the shapes and guess the letters. That guessing is what can go wrong, and it is why the second half of this page is about the photograph rather than about the software.

The picture matters more than the tool

Every recognition engine, cheap or expensive, does better on the same three things: straight, flat, and evenly lit. A photograph taken from an angle makes each line of text a slightly different height, so the engine is reading a page that curves away from it. A page held in one hand curls, and the curl does the same thing. Lying flat on a table under a ceiling light beats held up in the air every time.

Fill the frame with the page. A photograph where the text occupies a tenth of the image is giving the engine a tenth of the detail it could have, and cropping afterwards does not put back what the camera never recorded. Avoid the shadow of your own hand or phone — that is the single most common reason one half of a page reads perfectly and the other half comes back as nonsense — and avoid flash, which reflects off glossy paper and bleaches a stripe across the middle.

Do not zoom in. It feels like it should help and it does not: a zoomed photograph of six words has less context for the engine to work with than a whole page, and the letters are no larger in pixels than they were before. Put the whole page in the frame and let the software do the enlarging.

What to do with what comes back

Treat the output as a rough draft that saved you typing, and read it against the picture before you send it anywhere. Numbers are where mistakes hide: 1 and l, 0 and O, 5 and S, and the rn-versus-m confusion that has been the classic failure of every engine for thirty years. Names and amounts are worth a second look for the same reason.

The good news is that the mistakes are usually obvious in context, which means a spell checker and a quick read catch most of them. What they will not catch is a wrong digit in an account number, so anything financial is worth checking digit by digit against the original rather than trusting the whole.

Questions

Can I copy text from a screenshot?

Usually yes — but try selecting it first. Text in a web page, a message or a PDF can often be selected and copied exactly, with no guessing at all. Recognition is only needed when the screenshot is of something that was never text, or when the text cannot be selected.

Why did it read an 'm' as 'rn'?

Because at low resolution those shapes genuinely are almost identical, and this particular confusion is older than the software you are using — it is why old scanned newspapers are full of strange words. A sharper, closer photograph usually fixes it. Nothing else does.

Is a photo or a scan better?

A flatbed scan is better, every time, because the paper is held flat against glass and lit evenly. A phone is perfectly good if the page is flat on a table, the whole of it is in the frame, and your shadow is not falling across it.

Does it work on handwriting?

No, and it is worth being blunt about that. Handwriting recognition is a different and much harder problem than printed text, and the engines that attempt it are trained on thousands of hands rather than on yours. For a form filled in by hand, typing the answers is faster than fighting the software.

Is there a way to do this without any website at all?

Often, yes, and you should use it when it is there. Windows PowerToys includes a Text Extractor that reads the screen on a keyboard shortcut, and macOS has Live Text built into Preview and Photos. Both are faster than any web page, because the picture never has to travel anywhere. A website is for the machines and situations that do not have them — which is exactly why this one runs entirely in your browser rather than on somebody's server.

All tools