Extract text from scans and images

Document is a picture and you cannot select the text? This tool reads the letters out of the image itself and hands you text you can copy. It works in Arabic and English, and every step runs on your device.

Drop your files here

Files are processed on your device and never uploaded

Runs in your browser. Your file never leaves your device.

Take a photo

You can process several files at once, up to 3.

100% local, nothing uploaded

Three steps, that's it

Drop a scanned PDF or a photo
Pick the document language and a page range if you need one
Click extract, then copy or download the text

What does the OCR tool do?

When a document is a picture there is no text inside it at all: there are pixels drawing the shapes of letters, and nothing you can select, search or copy. Optical character recognition looks at the image and works out the letters from it, handing you text you can copy and search. This tool reads Arabic and English, and it takes a scanned PDF or a JPG, PNG or WebP photo straight off a phone camera.

The whole engine runs inside your browser, and we go further: the engine files and the language models are served from our own servers rather than a third-party network, so not one request goes to anybody else while your document is being read, and the file is never uploaded anywhere. That is the difference that actually matters when the page is a medical report, a bank statement or a signed contract.

When do you need Arabic OCR?

The first case is a document that reached you as a scan and you need a paragraph or a figure out of it: an old lease, a certificate, a letter from an office that prints and scans rather than exporting a text PDF. The second is a photo you took with your phone of a book page, a sign or an invoice, and you want its words rather than retyping them letter by letter.

The third is a whole archive: scanned files stacked on your device that your computer's search finds nothing in, because to it they are pictures rather than words. The searchable PDF option exists for exactly that. And the fourth is another tool turning your file away: PDF to Excel and extract text both come back empty on a scan, and the route out of both runs through here.

Text you copy, or a searchable PDF?

The first output is raw text: copy it with the copy button or download it as a text file, with a separator carrying each page's name when the file has more than one page. Pick it when you want the words themselves, to drop into a message, a document or a cell in a spreadsheet, and the original layout does not matter to you.

The second output is your file exactly as it was, with its pages, its images and its layout, and an invisible text layer over every page we read, so the document becomes searchable, selectable and copyable in any PDF reader without looking any different. We draw that layer ourselves, word by word from the recognition's own word boxes in logical order with an embedded Arabic font, because ready-made output from recognition engines writes Arabic in visual order and a search through it finds nothing.

How to extract text from a scan, step by step

Open the OCR page and drop a scanned PDF or an image into the tray, or tap the tray to pick the file from your device or from your phone's photo library.
Pick the document language: Arabic, English, or both together, which is the default and the right choice for a page carrying both.
Pick the output: text you can copy, or a searchable PDF that keeps your pages looking exactly as they do.
Type a page range if you only want part of the file, such as 1-3. That is also the way through when the file is longer than one run can read.
Press extract text. The first time, the recognition engine and the language model download once and then stay cached in your browser, so you never wait for that again.
Copy or download the text, and review the numbers, names and dates before relying on any of them in an official document.

Limits to know about

One run reads up to 20 pages. The ceiling is technical rather than commercial: recognition is far heavier on the device than anything else here. Read a long file in batches with the page range, or split it first with the split PDF tool.
Accuracy honestly varies: high on clean print, and low on handwriting, tilted phone photos and poor printing. Do not rely on a figure or a name in the output before you have read it with your own eyes.
Page layout is not preserved in the text output: a table lands as lines, two columns can interleave, and footnotes mix into the body. If the shape of the page matters, pick the searchable output instead of raw text.
Choosing a single language weakens recognition of the other one on the same page. A mixed page should be read with Arabic and English together, which is exactly why that is the default.
The text layer in a searchable PDF is placed only on the pages that were read. If you set a range, the whole document comes back with the layer on that range alone.
You can drop more than one file in a run, and a run takes up to 3 files. That ceiling is technical too, because every page is rendered and read inside your own browser.
A password-protected file will not open for reading. Unlock it first with the unlock PDF tool and the current password.

Common problems and fixes

The text came out scrambled or full of strange letters

The image is most likely tilted or unevenly lit. Run the file through the clean up scan tool first, which straightens the page and evens out the paper white even when half the photo is in shadow, then read the cleaned result.

No text came out and I got an error

The two messages are deliberately different. One says the recognition engine did not start, which is a download or connection problem and the tool retries with a fresh copy of the model on its own. The other says we read the pages and found no text, which means an image too weak to be read.

I need a Word file from a scanned document

Do not come through here. Drop the file straight into the PDF to Word tool: it offers to recognise the scan and builds the docx on that same page, instead of you copying text from here and pasting it there.

Search in the searchable PDF does not find an Arabic word

Check that you are searching the output file rather than the original, and that the page is inside the range you actually read. And if the word itself was misrecognised, no search will find it, which is a limit of the accuracy rather than of the search.

The first page is taking a very long time

The first download of the engine and the language model is a few megabytes. If it stops making progress altogether, the tool tells you the engine did not start rather than leaving you on an endless bar, and retries with a fresh copy. Try a better connection, and later runs will start immediately.

The file is large and my phone stalls while reading

Read it in batches with the page range, or shrink it first with the compress PDF tool. Recognition renders every page at twice its size before reading it, and that is the heaviest step on a phone's memory anywhere on this site.

I photographed a page with the camera and it read worse than a scan

A camera photo carries tilt, shadow and a curve in the paper. Shoot from directly above the page with even light and fill the frame with it, then run it through the clean up scan tool before reading.

Frequently asked questions

How do I extract Arabic text from a scan or a photo?

Drop the image or the scanned PDF into the tray and pick the document language: Arabic, English, or both together, which is the default. Then pick an output: text you can copy, or a searchable PDF. Press extract and wait a few seconds per page. The tool takes PDF files and JPG, PNG and WebP images, and all of the reading happens inside your browser.

How accurate is Arabic OCR?

High on clean printed Arabic or English, and noticeably lower on handwriting, tilted phone photos and badly printed pages. We render each page at twice its size before reading it, because Arabic diacritics and small print dissolve below that, but no amount of rendering makes a bad photo good. Always review the output before relying on it, especially numbers, names and dates.

Why is the first page slower than the rest?

The recognition engine and the language model download once on first use, a few megabytes, then stay cached in your browser so later runs start immediately. Recognition itself takes a few seconds per page on a phone. If the download stops making progress altogether, the tool tells you the engine did not start rather than leaving you on a progress bar, and retries with a fresh copy of the model.

Is my file uploaded to a server while it is read?

No. The recognition engine runs inside your browser, the file stays on your device and it never passes through us. We even host the engine and the language models on our own servers rather than a third-party network, so no request goes to anyone else while your document is being read. A statement, a medical report or a contract is read without being uploaded anywhere.

How is this different from the extract text tool?

The extract text tool copies text that already exists inside a text-based PDF, which is faster and far more accurate because it guesses at nothing. This tool is for the harder case: a document that is a photo or a scan with no text layer, where the letters have to be read out of the picture. Try extract text first; if it comes back empty, your file is a scan and this is the route.

What is the searchable PDF option and when should I pick it?

It returns your file looking exactly as it did, with an invisible text layer over every page we read, so the document becomes searchable, selectable and copyable in any PDF reader. Pick it when you want to archive the document as it is rather than lift its words out. We draw that layer ourselves from the word boxes in logical order, because ready-made output writes Arabic in visual order and search finds nothing.

How many pages can it read in one run?

Up to 20 pages per run, and that is a technical ceiling rather than a commercial one: recognition is far heavier per page than anything else here, and an unbounded run exhausts a phone's memory before it finishes. For a longer file, type a range in the pages box such as 1-20 then 21-40, or split the file first with the split PDF tool and read each part. No subscription lifts that number.