Pull the text out of a JPG photo without uploading it anywhere. Recognition runs in your browser, so the photograph of your passport or your payslip stays on your device. Worth knowing why JPGs are the hardest case: the format saves space by discarding fine detail, and the sharp edge of a letter is precisely the detail it discards first — which is where the faint halos around compressed text come from, and why an 'e' starts reading as a 'c'. On top of that, a JPG of text is nearly always a photo rather than a scan, so angle, shadow and focus are working against you too. No OCR engine recovers strokes that compression already removed, so the useful advice is about how you take the picture, and this page gives it.
| Problem | Effect |
|---|---|
| Heavy JPEG compression | Halos around letters; e reads as c |
| Photographed at an angle | Skewed lines the engine mis-segments |
| Shadow across the page | Part of the text drops below threshold |
| Text small in the frame | Too few pixels per character |
| Handwriting | Largely unsupported |
No. Recognition runs in your browser using Tesseract compiled to WebAssembly. Photos of documents are usually the most sensitive files people own, which is exactly why this should not touch a server.
Two reasons stack up. JPEG compression removes the high-frequency detail that makes letter edges sharp, and a photo adds perspective skew, uneven lighting and focus problems that a flatbed scan does not have. The recogniser is the same; the input is harder.
Get the page flat, light it evenly without a shadow across it, hold the camera square rather than at an angle, fill the frame with the text, and shoot at the highest quality your camera offers. Those choices matter far more than anything you can change after the fact.
For a screenshot, substantially — PNG is lossless, so screen text has no compression damage at all. For a photograph it makes little difference, because the camera has usually already encoded to JPEG before you see the file.
Several, selectable before you run it. Picking the right one matters: an English model reading French will guess accented characters wrong, because it does not expect them.
Not reliably. Tesseract is trained on printed type. Neat block capitals sometimes work; ordinary cursive does not.
JPEG is where OCR accuracy is usually lost, and it is lost before the file ever reaches a recogniser. The format discards high-frequency detail to save space, and the sharp black-to-white transition at the edge of a letter is exactly the high-frequency detail it throws away first — which is why heavily compressed text develops the faint halos and smeared edges that make an 'e' read as a 'c'. Recognition runs through Tesseract compiled to WebAssembly, entirely in your browser, but no engine recovers strokes the compression already removed. That matters here more than for other formats because JPEGs of text are almost always photographs rather than scans: a phone held over a page, at an angle, under uneven light. Perspective skew, shadow gradients across the page and the camera's own sharpening all compound the compression loss. The practical advice the page gives is therefore about capture rather than settings — flat, evenly lit, filling the frame, at the highest quality the camera offers — because those choices determine the result far more than anything done afterwards. Files are processed as a batch on your device, and the image never leaves it.