Image OCR & Text Extraction

BrainCoder's image OCR tool guesses the printed text in a picture and hands it back as plain text you can copy or download. Open a PNG, JPEG, GIF, WebP or BMP file, pick the language the text is written in, and press Extract — the recognizer runs as a WebAssembly build of Tesseract in your browser tab. Nothing is uploaded: the image, the recognizer and the language model are all read locally or served from this site's own origin, so a contract, a payslip, an ID or a whiteboard photo never reaches a server.

Twelve languages ship with the tool, each with its own model downloaded on that language's first use and then cached by your browser: English, Spanish, French, German, Portuguese, Italian, Russian, Hindi, Arabic, Chinese (simplified), Japanese and Korean. They are separate models, one per run — you cannot read a Danish page and its English header in a single pass, and each model is a real download you can see the size of before you start (English is about 3.0 MB gzipped, French about 0.7 MB, Japanese about 2.0 MB).

The limits are stated rather than hidden, and a file over any of them is refused with its actual numbers instead of being quietly cropped or shrunk: 25 MB per file, 16 megapixels per image and 8192 px on the long side. The format is confirmed from the file's own header bytes before the image is decoded, so a file that only claims to be a PNG is rejected instead of failing somewhere inside the recognizer. Output is cleaned up before you see it — CRLF and form-feed line breaks become plain newlines, runs of trailing spaces are trimmed and runs of blank lines are collapsed — and it is offered as clipboard text and as a .txt file named after your image.

What OCR is: a guess. Tesseract scores its own output, and this tool shows you that score with a plain description of what it means, because a number on its own tells you nothing. The engine is reliable on clean, straight, evenly lit printed text and it is unreliable on handwriting, stylised fonts, skew, shadows and low-contrast photographs. Two things happen to a very large photo that this tool does not control: the engine downsamples it before recognising, and the file's EXIF rotation flag is not applied, so a phone photo shot sideways comes out sideways. Scans at roughly 300 DPI and deskewed, evenly lit crops are the input this engine was built for — read every result over before you rely on it.