Image to text (OCR)

How to use it

1. Choose a clear image of printed text — a screenshot, photo or scan.
2. Press "Read text". The first run downloads the OCR engine once, then it is cached.
3. Copy the recognised text. Straight, sharp, printed text reads best.

Resolution decides more than the engine does

Recognition needs enough pixels per letter before it can separate an e from a c. The practical floor is about the equivalent of a 300 DPI scan, which for ordinary body text means lower-case letters standing roughly twenty pixels tall. A page photographed from across a desk fails for that reason alone. Move the camera closer or scan larger; enlarging a small photo first adds pixels but no new letter shapes.

Contrast and skew follow. Pale grey text on white, or a photograph of a lit screen with moiré banding across it, blunt the edges the engine looks for. Mild tilt is corrected automatically, but a steep angle, or an open book whose lines curve into the gutter, defeats line detection before recognition begins. Expect a few wrong characters even on clean input, and check reference codes by eye: nothing in the surrounding words flags a 0 read as O.

The scanned PDF is the case only OCR can solve

A PDF exported from a word processor holds real text objects, and an extractor reads them straight out. A PDF produced by a scanner or an office photocopier holds a photograph of each page and no text layer at all, which is why the PDF to Text tool returns nothing for one. Export the page as an image and run it here; recognition is the only route back to characters.

Layout is where output most often surprises people. The engine infers a reading order, so a two-column page can come back with a line from column one followed by a line from column two, alternating downwards. Tables lose their alignment for the same reason: the recognised words are right, their arrangement is not. For anything columnar, crop to a single column, read it, then crop the next.

Questions