A screenshot of a paragraph you want to quote. A photo of a receipt you need to expense. A scanned page you'd rather not retype word for word. The text is right there in front of you, but it's locked inside a picture — and copying it by hand is exactly the kind of tedious job worth handing to a computer.
What OCR actually does
OCR — optical character recognition — is the technology that turns the shapes in an image back into editable characters. It looks at the strokes that make up each letter, compares them against a trained model of what letters look like, and reconstructs the words. The engines behind the good free tools (Tesseract is the well-known one) have been trained on enormous amounts of text and are genuinely excellent on clean, printed type. They're far less sure of themselves on anything messy.
Better results come from a better image, not a better tool
This is the part people skip. The single biggest factor in OCR accuracy is the picture you feed it. A few things make a real difference:
- Resolution. The text should be crisp, not a tiny smudge. For a scan, aim for around 300 DPI. For a photo, get close enough that the letters are sharp.
- Contrast. Dark text on a light background is ideal. Faint grey text or text over a photo confuses it.
- Straightness. A page shot at an angle, or rotated, throws it off. Square it up first if you can.
- Crop tight. Trim away the desk, the margins, the other columns — give it just the text you want.
- Pick the right language before you run it. Telling it the text is French produces very different (and better) results than leaving it on English.
What OCR still struggles with
It's worth being honest about the limits so you're not surprised:
- Handwriting — especially cursive — is unreliable. These engines are built for printed text.
- Decorative, script, or heavily stylised fonts trip it up.
- Blurry, low-resolution, or heavily compressed images lose the detail it needs.
- Text laid over a busy background or a photo is hit and miss.
How to extract text from an image
- Open the Image to Text tool and drop your image in.
- Choose the language of the text before it runs.
- Let it work. The first time, it downloads the language model — a few megabytes — then caches it, so the next run is quicker.
- Copy the result or download it as a .txt file, and give it a quick proofread.
That first-run download is why OCR always feels a little slow the very first time and snappy afterwards. It's also why this runs privately: once the model is cached, the recognition happens on your device, and the image never leaves it.