How to get the text out of a picture or a scan, without uploading it

Drop the picture, the screenshot or the scanned PDF into OPTIC, pick the language it is written in, and read. The recognition engine runs inside the tab — the first time, it and the language file are fetched once from this site and kept on your device, after which the page works offline. The text comes out to copy or save, every word is boxed over the picture so you can see what came from where, and a scanned PDF can be saved back as a PDF whose text is searchable. The picture is never uploaded.

What OCR actually is

A photograph of a page is a grid of coloured dots. Nothing in it says "this is the letter e". Optical character recognition finds the lines, the words and the letter shapes, and matches each against a model of what letters look like in that language. It is pattern recognition, and it is very good at the thing it was trained on — printed text, roughly square-on, in a font a book might use — and progressively worse as the picture moves away from that.

There is no threshold where it stops working and tells you. It degrades: a nine becomes a four, an rn becomes an m, a column of figures loses its order. Which is why the words are drawn back over the picture here, and why a confidence figure is shown for the page. If it says 94% you can skim it. If it says 61% you should read every line against the original.

What it reads well

• A scan or a screenshot of printed text. This is the case it was built for, and it is close to perfect.
• A photograph of a page, taken square-on, in even light, in focus, filling the frame.
• A scanned PDF, page after page. Pages that already hold real text are read as text rather than guessed at.
• A phone photograph in HEIC, without converting it first.

What it reads badly, or not at all

This is the part the sites with a big Upload button leave out.

• Handwriting. Not supported, at all, and no setting will help. The engine models printed letterforms; handwriting is a different problem needing a different kind of model.
• Photographs taken at an angle. Lines that curve or converge break the line-finding before letter recognition is even reached. Square-on matters more than resolution.
• Small or blurry pictures. A letter needs roughly twenty pixels of height to be read reliably. A screenshot of a screenshot rarely has it.
• Columns, tables and forms. The letters come out; the reading order may not. Two columns can interleave into one line each. Expect to fix the layout by hand.
• Decorative, condensed or very old type. Blackletter, heavy italics, dot-matrix and worn fax print are all much worse than clean body text.
• The wrong language. Choosing English for a French document costs you every accented word. The language chooser is not a formality.
• Text over a photograph, low-contrast text, and white text on dark backgrounds, which vary from fine to hopeless.

Getting a better read

Most bad results are bad pictures. Before blaming the engine: take the photograph square-on rather than leaning over the page; get more light rather than more pixels; fill the frame with the page; and if the picture is skewed, straighten it in PHOTO first — a two-degree rotation can be the difference between 60% and 95%. If a scan is very dark or very faint, raising the contrast helps more than any OCR setting.

Making a scanned PDF searchable

A scanned contract is a stack of pictures. You cannot search it, cannot copy a clause out of it, and a screen reader finds nothing in it at all. OPTIC reads each page and saves a PDF that looks exactly the same but carries the recognised text behind the picture, so Ctrl-F finds a clause and a reader can speak it. The picture stays the authoritative thing on the page; the text sits under it.

That is also the right order if the file has to get smaller: read it first, then compress it with PRESS, so what you end up with is small and searchable rather than small and mute.

Why it matters that it stays here

The documents people run through OCR are the ones they cannot retype: a payslip, a bank statement, a medical letter, a passport page, a contract, somebody else's correspondence. Every free OCR site takes a copy of that file onto a machine you do not control, and the promise that it will be deleted in an hour is a promise about housekeeping, not about what has already happened. Here the file is read in the tab. Load the page once, turn the network off, and it still works — which is the only form of that promise you can check yourself.

The short version