Copy text out of an image or a scanned PDF

What happens when you read a page in

In a photograph or a scanned PDF the text is not text, it is pixels. You cannot search it, you cannot copy from it, and screen readers can do nothing with it. Text recognition — OCR — translates that back into letters. It takes two steps, and both of them run here, in your browser. The first is a neural network that finds where there are lines on the page at all: it looks for strokes rather than letters, which is why it finds them on a crookedly scanned page or one photographed with a phone. The second is another network that reads what the first one found, line by line, and gives back characters.

The models download on first use — roughly 25 megabytes, runtime included — and then stay in your browser's cache, so the next page starts instantly. The file never reaches a server in the meantime. That is not a marketing line; it is also why a page takes seconds to process: your own machine is doing the arithmetic, not a server farm.

What makes a scanned PDF searchable

A searchable PDF is not a new document: it is the same page image you scanned, with the difference that the recognised text is placed over the words as well — invisibly, exactly where the ink is. Nothing changes on screen. Ctrl+F, on the other hand, finds the word you are looking for, the text can be selected and copied, and a screen reader can read it aloud. That is why authorities and grant applications often ask for scanned paperwork as a „searchable PDF".

The page image goes into the searchable PDF re-compressed, as JPEG. That is deliberate: the text layer's coordinates belong to exactly the image recognition saw — smuggling the original file back in would embed some photos rotated, and the selection would land beside the words. The size is therefore usually similar to the original, but not identical to it.

Where its limits are

Honestly: it reads printed text, not handwriting. It knows the Hungarian accents, but the two long double-acute vowels sometimes come back at small type sizes wearing a single acute or an umlaut instead. It is worth skimming the result before you send it anywhere. Table structure does not survive either: you get the contents of the cells, not the grid.

What matters a great deal is resolution. An A4 scanned at 200 DPI, or a phone photo taken straight on in good light, is close to flawless; a crooked, blurry, shadowed image will defeat any OCR. One file can be at most 40.00 MB, and from a PDF we read the first 20 pages — every page is a separate run, and a hundred-page bundle would take so long that nobody would wait for it.

Frequently asked questions