Image to Text (OCR)

Pull the text out of a screenshot, scan or photo — read on your device.

Processed locally — nothing is uploaded

About Image to Text (OCR)

Tesseract — the recognition engine behind most open-source OCR — compiled to WebAssembly and served from this site rather than a CDN. The engine travels to your image instead of your image travelling to a server, which is also why the tool keeps working with the network switched off once it has loaded.

Drop in a screenshot, a scanned page or a photo of a document and get the text back as characters you can edit, search and paste. The reading is done by Tesseract, which began at Hewlett-Packard in the 1980s, was released as open source in 2005 and spent the following decade under Google's stewardship — it remains the most widely deployed engine of its kind. The picture goes into memory, gets read, and goes no further.

That last point is the whole reason this tool exists in this form. Think about what people actually put through an OCR service: a passport page for a visa application, a payslip for a mortgage broker, a doctor's letter, a contract with a rate in it, a utility bill as proof of address. Uploading any of those to a free web converter means handing a stranger's server a document you would not email to a colleague, and the privacy policies of that category of site are rarely reassuring reading. Here the engine comes to the image rather than the image going to a server.

Every word comes back with a confidence score, and the tool shows you the summary rather than hiding it. This matters because OCR fails quietly: it does not leave a blank where it could not read something, it produces a plausible wrong word. 'rn' becomes 'm', '0' becomes 'O', '1' becomes 'l', and the result reads perfectly until someone acts on a wrong account number. The verdict above the text is deliberately strict — a page counts as clean only when the engine was unsure about nothing at all, because as soon as there is a doubtful word, telling you to look at it beats telling you the page is fine.

Accuracy is mostly decided before the image reaches the engine, and a few habits change the outcome more than any setting. Aim for around 300 DPI: a scan at 150 loses the fine strokes that distinguish similar letters, while going far above 600 adds paper texture rather than detail. Keep the page square to the frame — Tesseract corrects small skews and struggles past about ten degrees. Photograph documents flat with even light rather than at an angle under a desk lamp, because a shadow gradient across a page defeats the thresholding step that separates ink from paper. And crop to the text you want before reading, which both speeds up the pass and stops the engine trying to make words out of a letterhead graphic.

The layout setting is the one control worth understanding. Tesseract analyses page structure before reading, deciding what is a column, what is a heading and what order to read blocks in — and that analysis is where results most often go wrong. Whole page suits a normal document. Single block is for one paragraph with no other layout. Column of text handles receipts and narrow invoices, where the default can interleave two columns into nonsense. Single line is for a caption or one field of a form. Scattered text turns off layout analysis for signs, screenshots and anything where the text has no reading order at all. If a result comes back scrambled rather than misspelled, the layout mode is almost always the fix, not the image.

One honest limitation: this reads printed text, not handwriting. Handwriting recognition is a different problem, needing different models, and Tesseract's accuracy on cursive is poor enough that offering it would be a disservice. English only for now, too — each additional language is another model to download, and shipping one that works well beats shipping ten that half work.

Learn how this works

Frequently asked questions

Is my image uploaded to a server?

No. The OCR engine is downloaded to your browser and the image is read there. This is verifiable rather than a promise: open your browser's network tab, read an image, and you will see the engine and language model being fetched from this site and nothing being sent back. After the first read, the tool works with your network disconnected.

Why is there a download the first time?

The engine is roughly 4 MB — a WebAssembly build of Tesseract plus the English language model. That is the price of doing the work on your device instead of a server, and it happens once per browser: it is cached afterwards, so later images start reading immediately. A server-side OCR service skips this download by having your image instead.

Can it read handwriting?

No, and it will produce confident nonsense if you try. Tesseract is trained on printed and typed text. Handwriting recognition needs different models entirely, and the accuracy on cursive is low enough that including it would waste your time rather than save it. Neat block capitals sometimes work; joined-up writing does not.

The text came out scrambled or in the wrong order. What went wrong?

That is almost always page-layout analysis rather than character recognition. Try the layout selector: 'Column of text' for receipts and narrow invoices, 'Scattered text' for screenshots and signs with no reading order, 'Single line' for one field. Scrambled word ORDER points at layout; misspelled WORDS point at image quality.

Can it read a PDF?

Not directly — this reads images. For a scanned PDF, export or screenshot the page you need as an image and read that. Note that many PDFs already contain real text and need no OCR at all: if you can select the text in a PDF viewer, copying it directly is both faster and perfectly accurate.

What does the confidence percentage actually mean?

It is the engine's own estimate of how sure it is about each word, averaged across the page. It is a useful signal, not a guarantee — a confident misread is entirely possible, particularly with digits, which have no dictionary to check against. Treat anything involving numbers as needing a human read regardless of the score.

Pro Tips

  • Crop to the text you care about before reading: it is faster, and it stops the engine trying to find words in logos and page furniture.
  • Scan at 300 DPI. Below 200 the strokes that separate similar characters disappear; above 600 you are mostly capturing paper grain.
  • If word ORDER is wrong, change the layout mode; if the WORDS are wrong, improve the image. They are different failures with different fixes.
  • Straighten the page before reading — Tesseract handles a few degrees of skew and degrades sharply past about ten.
  • Always proof-read numbers by hand. Digits have no dictionary behind them, so they are where confident misreads concentrate.
  • For a screenshot, capture at full resolution rather than resizing down first — an enlarged small screenshot reads worse than the original.

Common Use Cases

  • Pulling a quote or a paragraph out of a screenshot without retyping it.
  • Getting the figures off a scanned invoice or receipt for expenses, without sending the document to a third party.
  • Extracting text from an image-only PDF page that refuses to let you select anything.
  • Digitising a printed page — a recipe, a form, a letter — into something searchable.
  • Reading text out of a photo taken at a conference slide, a noticeboard or a shop sign.
  • Handling documents that carry personal data, where uploading to a free converter is not an acceptable option.

How It Compares

Online OCR services like OnlineOCR.net, i2OCR and the free tiers of the big PDF suites all work by uploading your document to their servers, where retention terms and staff access are things you take on trust. Google Docs and OneDrive can OCR an uploaded image and do it well, but only inside their accounts. Adobe Acrobat's OCR is excellent and priced accordingly. This tool trades away multi-language support and handwriting for the one thing none of those can offer: the document does not leave the machine, and the tool keeps working offline once loaded.

Related tools