Scanned PDF to Text (OCR)

Getting a cleaner read

Every page is rendered at 200 DPI before it is read, so the limit is the scan itself. Body text scanned at 300 DPI reads well; a phone photo of a page taken at an angle, or a fax-quality scan of small print, produces guesses. Pages that are sideways or upside down read very badly, so turn them upright with rotate PDF before running the recognition.

Pick the language the document is written in, since each one is a separate model. Chinese comes in Simplified and Traditional models, and choosing the wrong one costs a lot of accuracy. A document mixing two languages reads best under the one used for most of the text.

Checking what came out

The confidence figure is an average across all pages, so one bad page can hide inside a good score. Around 90 or above usually means only occasional slips; below 75, expect errors in most paragraphs. The usual confusions are 0 and O, 1, l, and I, rn read as m, and 5 as S, which matters most in amounts, dates, and reference numbers, so check every figure against the page.

Leave Join lines into paragraphs on for prose, where it mends sentences broken at the end of each printed line. Turn it off for addresses, lists, and tables, where the line breaks carry meaning.

How to use it

1. Choose a scanned PDF.
2. Pick the language of the text.
3. Read the result and copy or download it.

Privacy & limitations

Your file is never uploaded. The recognition engine and language model are downloaded once — around 10 MB — and cached by your browser, then everything runs on your own machine. Hosted OCR services work the other way round: they want the document. A long document takes a while, because each page is rendered and then read in turn.