FilesShifters

Turn a scan into text you can copy and search

It runs in your browser: your files stay on your device and are never uploaded.

Experimental It uses a model that runs in your browser, it can get things wrong, and it has not yet been tried on real phones.

Drop a scanned PDF or photos here, or press to choose them

A scanned PDF, or JPG, PNG or WebP images

A contract that went through a scanner, a photo of a page, an old PDF made only of pictures: with OCR the words become text you can copy and search.

How to recognize the text in a scan

  1. Add the scan. A PDF made of images, or one or more photos in JPG, PNG or WebP: the order of the cards is then the order of the pages.
  2. Pick the language. Italian, English or both: the engine knows the words of the language you give it, and makes more mistakes on text in the other one.
  3. See what will be downloaded. Before anything is fetched, the panel tells you how much the engine and the language data weigh, or that they are already on this device.
  4. Press “Recognize text”. The engine arrives first if it is missing, with its own bar; then the pages are read one at a time, with another bar.
  5. Check it and download. You get a text file and, if you left it on, the searchable PDF. The note under the result points to the pages worth reading more carefully.

What gets downloaded, and why

Reading letters in a picture takes a recognition engine and the data for the language: 2.9 MB for the engine, 1.6 MB for Italian, 2.8 MB for English. They come from this site, once, and only after you press the button.

That makes 4.5 MB for Italian, 5.7 MB for English and 7.3 MB for both. It is more than the rest of the page put together, which is why it does not come with the page: someone who lands here for another reason, or only wants to read how it works, does not pay for it. Hovering over the tool does not start it either.

The engine is tesseract.js, the browser build of Tesseract, a free text-recognition program that has been in use for decades; the language data is the compact kind, made to weigh little for the same result. Everything is served from filesshifters.com, under /ocr/, together with the licences: no other site receives a request. After the first time the files stay in the browser’s cache and the panel says so; with the page open, it works without a connection too.

The searchable PDF

It is your document with the recognized words underneath, invisible, each where it appears in the image: in a PDF reader you search for a word and find it highlighted on the scan, you select a line and copy it.

The pages are not redrawn or recompressed: the scanned images stay the same bytes, and the text layer adds only what is needed to search it. If you start from photos, each photo becomes a page as large as the photo, as in JPG to PDF; WebP images become PNG, losslessly, because a PDF cannot hold them.

The layer uses the standard PDF typeface, which covers Italian and English with every accent. A character from another alphabet, should the engine find one, would be left out of the layer (the note says in how many words) but not out of the text file.

How far to trust the result

Recognition makes mistakes: an “l” read as “1”, a lost accent, a split word. The engine reports how sure it is of each word, and the note under the result lists the pages with too many uncertain ones.

We do not give you an accuracy percentage, because the engine does not know how wrong it was: it only knows how sure it is. A clean page, printed and scanned straight, almost always comes out right; names, account numbers, amounts and dates are worth checking anyway, because those mistakes cost the most. The pages of a PDF are read at 200 dots per inch: we tried 150, 200 and 300 on the same scans, and 200 made fewer mistakes than 150 and as few as 300 (fewer, on the hardest one), in two thirds of the time.

What this tool does not do

  • It does not read handwriting. The engine is trained on printed type.
  • It does not rebuild tables. Cells come out as lines of text, in the order the engine finds them.
  • It does not read formulas. Fractions, subscripts and symbols come out as scattered characters.
  • It does not straighten pages. A crooked page, or one photographed at an angle, reads badly: better to retake the photo from above, with the page filling the frame.
  • It only knows Italian and English. Other languages with the same alphabet come out with more mistakes; other alphabets do not come out.
  • It does not re-read pages that already have text. It skips them and says so: that text comes out, with no reading mistakes, with PDF to text.
  • It does not take documents that are too large. At most 50 pages to recognize in one go, and 40 megapixels per page: beyond that the button stays off and the panel says why, before anything is downloaded. Split a long PDF with Split PDF, shrink a huge photo with Resize images.
  • It does not open a password-protected PDF. Unlock it first with Unlock PDF, if you know the password.

How to check for yourself that nothing leaves the browser

Open your browser’s developer tools, Network tab, and add the scan: no request for the recognition goes out. When you press the button the files under /ocr/ appear, all from this site, and no request carrying your document.

The second test is to cut the network. After the first time, with the page still open, take the computer or phone offline and recognize another page: the result still arrives, because everything it needs is already in the browser.

Frequently asked questions

Is the scan uploaded anywhere?

No. The pages are read in your browser, on your device, and the text is created there. From the network come only the engine and the language data, the first time; nothing goes out to it.

Why do I have to download something?

Because recognition needs an engine and language data that weigh a few megabytes. Instead of sending your file to a server that has them, the engine is brought to you. It downloads once, from this site, and only when you press the button.

How long does it take?

On a recent computer an A4 page takes a couple of seconds, plus the time to download the engine the first time. On a phone it is slower, and with many pages it helps to keep the screen on; you can cancel at any moment.

Does it work without a connection?

Yes, after the first time and with the page open: the engine and the language stay in the browser. The page itself has to be opened with a connection, because the site does not keep its pages offline.

How is this different from PDF to text?

PDF to text reads the text a PDF already contains, and it is exact. This tool reads the words in a picture, where there is no text: it is for scans and photos, and it can make mistakes.

How can I check that this is true?

Open the developer tools, Network tab, and recognize a page: you will see this page’s files, those of the engine under /ocr/, the requests of the visit counter described at the bottom of the page, and no request that carries your document.