Image to Text (OCR)

about this tool

Tesseract, on your device

Drop a photo, scan or screenshot and Tesseract — the open-source OCR engine that has been reading text since the 1980s, compiled to WebAssembly — reads it in a worker thread in your browser. Nothing is uploaded; this site has no server that could take the image. The first run fetches the engine (about 7 MB: the worker and the core build your browser's features select) and the language data you have ticked, all from this site, and the browser's ordinary cache keeps them. Nothing is written to IndexedDB or local storage.

Twelve languages are served here, using Tesseract's compact "best" models: English, Spanish, French, German, Portuguese, Italian, Dutch, Hindi, Arabic, Russian, simplified Chinese and Japanese. Tick more than one for mixed text. Tesseract has data for over a hundred further languages, but the usual source is a third-party server, and the privacy page says a page here makes no request to anyone else — so the list is what can be served first-party.

What comes back

The recognised text, tidied the way you would before pasting it: stray spaces collapsed, form feeds turned into paragraph breaks, words the scan hyphenated across a line break rejoined (only before a lower-case letter, so "Anglo-Saxon" survives), and the lines of each paragraph joined into one — both tidyings are switches. The text box is editable; edits are kept until you clear.

Tesseract's confidence is averaged over the words it found and turned into plain advice, and the words under 60% are listed so you can check them against the image. Exports: the text as .txt, a TSV of lines with confidence and bounding boxes in image pixels, and a JSON file with every word — for anyone building on the result.

Getting good results

Tesseract wants text at least 20 pixels tall, straight, in focus and dark on light, in the language you ticked. Screenshots and flat scans do well. Photographs of curved pages, decorative fonts, handwriting and low light do not; crop to the text with the resize tool, or rotate it, before reading.

Limits

Images only — render a PDF page to PNG first. Handwriting is largely beyond the engine. The language list is the twelve above. Recognition is single-threaded WebAssembly: a phone photo takes a few seconds, a dense A4 scan longer. Nothing is stored between visits beyond the browser's cache of the engine and language files.

For a PDF that already has a text layer, PDF to text is faster and exact — try it first. Recognition improves on a clean, straight image, so image resize & crop is worth a pass on a photographed page.

questions

Is the image uploaded to a server?

No. Tesseract runs as WebAssembly inside your browser, in a worker thread. The engine (about 7 MB) and each language you tick (one to three MB) are fetched from this site the first time and kept by the browser’s cache; the image itself never leaves the page, and this site has no server that could receive it.

Which languages are supported?

Twelve are served from this site: English, Spanish, French, German, Portuguese, Italian, Dutch, Hindi, Arabic, Russian, simplified Chinese and Japanese, and you can tick several for mixed text. Tesseract has data for over a hundred more, but those would have to be fetched from a third party, which this site does not do.

How do I get better results?

Give the engine what it likes: text at least 20 pixels tall, straight, in focus, dark on light, with the right language ticked. Crop away everything that is not text with the resize tool first. Screenshots do well; photographs of curved pages, handwriting, decorative fonts and low light do not.

What does the confidence figure mean?

Tesseract’s own estimate, averaged over the words it found: above 90% is usually correct as-is, 75–90% needs a read-through, and below that expect errors. Words under 60% are listed separately so you can check them against the image, and the text box is editable.

Can it read PDFs or handwriting?

Not here. A PDF has to be rendered to an image first — export the page as PNG and drop that. Handwriting is outside what Tesseract was trained on; results are poor except for very neat block capitals.