How to extract text from a photo or scan (free OCR)
Published September 26, 2026
A photo of a document contains text you can read but not use: you cannot search it, copy a paragraph, or fix a typo. Optical character recognition (OCR) closes that gap — software reads the pixels and reconstructs the characters as real, editable text. Here is how it works, how to get accurate results from your own documents, and where OCR still stumbles.
What OCR actually does
OCR examines the image, finds the regions that look like lines of text, segments them into characters, and matches each shape against what it has learned letters look like. Modern engines read both a crisp laser print and a decent phone photo — but they are only as good as the image. The cleaner the scan, the better the text: high contrast, straight lines, even lighting.
This is why running OCR on a corrected scan works so much better than on the raw photo. Perspective correction makes the lines straight, and document-mode enhancement makes the letters dark on a white page — exactly the input OCR engines were built for.
Extracting text in Scandut
Scan your pages as usual, then open the Extract text (OCR) panel on the results screen. One click reads every scanned page — Korean and English, including mixed text — and shows the recognized content in an editable box. Recognition runs entirely in your browser: the first run downloads the language data once (about 15 MB), and your document itself is never uploaded anywhere.
Read the result over and fix anything the engine misread — a zero for the letter O, a garbled word where a stamp overlapped text. Then download the text as a plain TXT file, a Word document, or an Excel spreadsheet where each line is split into columns.
Getting accuracy up
Three factors dominate accuracy. Resolution: text should be reasonably large in the frame — shoot closer, not wider. Contrast: use Document mode; gray text on gray background is the classic OCR killer. Orientation: rotate pages upright before extracting; OCR reads sideways text poorly. With a clean, corrected scan of printed text, expect near-perfect results; with a dim photo of small print, expect to do some editing.
What OCR is bad at
Handwriting is the big one — cursive and quick notes are still largely beyond reliable OCR, though neat block letters sometimes work. Ornate fonts, dense tables with tiny print, text over photos or watermarks, and very low-resolution images all degrade results. For anything where an error would matter — an account number, a legal clause — treat OCR output as a draft and verify against the image.
FAQ
Which languages are supported? Korean and English, including documents that mix both. Is my document uploaded for recognition? No — recognition runs in your browser; only the language model is downloaded, once. Why is the first run slow? That one-time download. Can it read a whole multi-page batch? Yes, all scanned pages are read in one run, and the export contains every page.