How to copy text from an image or screenshot

Text in a screenshot, a photo of a notice or a scanned letter is just a picture of letters — you cannot select it. Optical character recognition (OCR) reads the shapes and turns them back into text you can copy and edit.

Copy text from an image step by step

1. Open Image to Text (OCR) and add the image.
2. Choose the language of the text. This matters more than people expect: the engine uses a dictionary for that language to settle ambiguous shapes.
3. Press Extract text. The first run downloads the language data, then it is cached.
4. Check the result, then copy it or download it as a text file.

The recognition runs on your own device, so the image is not uploaded anywhere.

How OCR reads a picture

It helps to know what the engine is doing, because every tip below follows from it. It first looks for blocks that seem to contain text, then cuts each block into lines, and each line into individual character shapes. Each shape is compared with a model trained on printed type in the language you chose, and a dictionary helps decide between look-alikes. Anything that makes the lines hard to find, or the shapes hard to separate, costs accuracy — which is why a clean, straight, close-up image matters so much.

How to get clean results

Accuracy depends almost entirely on the picture you give it:

• Use the sharpest version you have. A screenshot beats a photo of a screen; a flat scan beats a photo of a page.
• Make the letters big enough. Tiny text is the most common cause of gibberish. As a rule of thumb, capital letters should be at least 20–30 pixels tall. If the text is small on screen, zoom in before you take the screenshot — enlarging a small image afterwards does not bring back detail.
• Straighten it. Text should be horizontal — fix sideways images with Rotate Image.
• Crop away the clutter. Backgrounds, logos and photos confuse the engine. Trim to just the text with Crop Image.
• Light it evenly. Shadows across a page and glare on glossy paper both cost accuracy.
• Prefer dark text on a light background. High contrast reads best. Pale grey text, coloured backgrounds and text laid over a photograph all reduce it.

On clear printed text — screenshots, rendered documents, flat scans — accuracy is typically above 95%. Still, proofread numbers, names and anything you will rely on.

Photographing a page properly

When a photo is the only option, a few seconds of care make a large difference:

• Lay the page flat and hold the phone directly above it, parallel to the paper.
• Fill the frame with the text rather than the table around it.
• Stand so your own shadow does not fall across the page; daylight from a window is ideal.
• Tap the screen to focus on the text, hold still, and take two shots to be safe.
• Turn the flash off on glossy paper — it leaves a white patch where the text used to be.

Columns, tables and mixed layouts

OCR returns text line by line, so layout is where it gets confused. A two-column page may come out with the columns interleaved, reading straight across the gap. Crop each column into its own image and run them one at a time. Tables lose their grid in the same way: you get the words and figures, but not which cell they were in, so expect to rebuild the columns by hand in a spreadsheet.

What to proofread

Errors cluster in predictable places. Check these first:

• Look-alike characters: 0 and O; 1, l and I; 5 and S; and the pair “rn” read as “m”.
• Numbers — account numbers, prices, dates and phone numbers, where a single wrong digit matters and a dictionary cannot help.
• Names, email addresses and web addresses, for the same reason.
• Punctuation and symbols — full stops lost, commas read as full stops, currency signs swapped.

Tidying the extracted text

The result keeps the line breaks of the original image, so a paragraph arrives as a stack of short lines. Paste it into Text Cleaner and switch on the option that joins broken lines: single breaks are merged into flowing paragraphs, and blank lines between paragraphs are kept.

What OCR struggles with

Handwriting. The engine is trained on printed type, so neat block capitals sometimes work and ordinary joined-up writing rarely does. Unusual decorative fonts and text set at an angle are also hit and miss.

Languages

Pick the language the text is actually written in, not the language of your phone or browser. The supported set includes English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean, Arabic and Hindi, and each one downloads its own trained model the first time you use it. For a document that mixes two languages, choose the one that makes up most of the text and proofread the rest.

When your device can already do it

For grabbing a line or two, check what is built in: recent iPhones and Macs let you select text directly in a photo, Google Lens does the same on Android, and the Windows 11 Snipping Tool can copy text from a capture. They are the fastest route for a phone number or a Wi-Fi password. A dedicated OCR page is more useful when you are on a computer without those features, need a particular language, or want the result as a text file.

Scanned PDFs

A scanned PDF is a set of page images with no text inside. Convert its pages to images with PDF to PNG, then run each page through OCR. See how to convert a PDF to Word for PDFs that do contain real text.