Image Text Extraction
Read the text out of a picture and get it back as copyable text
Loadingโฆ
Was this tool helpful?
About this tool
Read the text in a photo or screenshot and get it back as text you can copy. Korean and English are recognised together, and the result can be edited in place, copied, or saved as a TXT file. Recognition runs inside your browser through WebAssembly, so your image is never uploaded - only the language training data is fetched from a CDN, and only the first time.
How to use
- Drop an image in, paste one with Ctrl+V, or use the screen capture button.
- Choose a language. If the text mixes scripts, Korean + English works well.
- If the text is faint, turn on Boost contrast and press Read the text.
- Check the result, fix anything that came out wrong, then copy it or save it as TXT.
Frequently asked questions
- The accuracy is terrible.
- Resolution matters most. Recognition falls off sharply once letters drop below about 20 pixels tall, so retake the shot closer or zoom in before capturing. Skew is next: straightening a tilted page often helps more than any option here.
- When should I use grayscale or contrast?
- They help most on photos of documents lit from one side, where a broad gray cast sits over the page. Raising contrast pushes that gray to white and leaves only the letters. On an already crisp screenshot they do nothing, and they can break up thin lettering.
- Does it read handwriting?
- Barely. The recogniser is trained on printed text, so even neat block handwriting comes out with a much lower accuracy.
- Can I feed it a PDF?
- Not directly. Split out the pages you need with the PDF tool, then capture that view and drop it here. If the PDF contains real text rather than a scan, copying it straight from a viewer is far more accurate.
- Is my image uploaded anywhere?
- No. Recognition happens inside your browser and the image never leaves this tab. The only outbound request is for the language training data, and that happens once.
The training data is close to 20MB for Korean alone, so the first run takes a while. On a metered connection, do that first run on Wi-Fi. Always check the output yourself - digits and symbols in particular are often misread as similar-looking characters.