How to get the text out of a picture
- Add your pictures. Drop one or many anywhere on the page, paste one (a screenshot works), or click the box to choose files. Phone photos, including iPhone HEIC, work too.
- Choose the language of the text. If a picture has two languages, pick the second one under Also read.
- Press Read the text. Each picture gets a box with its text, the number of words, and how sure the reader was.
- Copy it or save it. Use Copy and Download .txt under a picture, or Copy all the text and Download as .txt at the top for everything at once.
The first time you press the button, the page loads the reading engine and the language data. That is about 4 MB for the engine plus 0.7 to 2.9 MB for the language (English is 2.8 MB), and you see its progress. Your browser keeps the language data, so on a later visit only the engine is needed again. After that a normal page of text takes a second or two.
The settings
| Setting | What it does |
|---|---|
| Language of the text | The most important setting. The reader uses a model for one language, so the wrong choice gives wrong letters. In our test, a Spanish sentence read with English chosen had 14% wrong characters, because the accents and the ñ were lost, and with Spanish chosen it had none. |
| Also read | A second language, for a picture that mixes two, such as English and Spanish. It is a little slower, and it is best left off when one language is enough. |
| The picture is | How the reader looks for text. Automatic finds columns and paragraphs by itself and suits pages of text. One block of text suits a single paragraph or a screenshot of a message. One line suits a caption, a label or a serial number. Scattered text suits words all over a sign, a poster or a diagram. |
| Turn the picture upright if needed | On by default. If the first reading is shaky, the page tries the picture turned a quarter, a half and three quarters, and keeps a turn only when it reads clearly better. The result says turned 180° to read it when it did. |
| Light text on a dark background | Flips the picture's colors before reading. The reader often copes with light text on dark by itself (our test picture was read perfectly either way), so tick this only when such a picture comes out wrong. |
| Join the lines of each paragraph | Turns the lines of a paragraph into one flowing line, so the text can be pasted into an email or a document without hard line breaks. A word split with a hyphen at the end of a line is put back together. It only changes what is shown, so you can switch it on and off without reading the picture again. Text that you have edited is left alone. |
If you change the language, the picture type or the two checkboxes that change how the picture is read, the finished texts are cleared, because they no longer match, and you press Read the text again.
What you get for each picture
- The text, in a box you can edit. Lines stay as lines, with a blank line between paragraphs.
- The number of words and the confidence. The confidence is the reader's own average certainty over the words, from 0 to 100. Above 85 the text is usually very good, from 65 to 85 check it against the picture, and below 65 the page shows a warning. It is the reader's opinion of itself, not a guarantee.
- How long it took, in seconds.
- A warning when nothing was found, so a picture with no readable text does not leave a silent empty box.
How well does it read?
We tested the reader on pictures that we made from text we knew, and counted the wrong characters. These are clean, computer-made pictures, so they show what the reader can do at its best:
| Picture | Result |
|---|---|
| Three paragraphs of text with numbers and a price, 900 pixels wide | No wrong characters, 95% confidence, 0.4 seconds |
| A full A4 page, 2480 × 3508 pixels, 44 lines, 396 words | No wrong characters, 2 seconds |
| Small screenshot text, 13 pixels high | No wrong characters (the page enlarges it first) |
| Italic serif text on cream, and grey text on a grey background | No wrong characters |
| Russian, French, German, Arabic, Hindi and Japanese, two lines each | No wrong characters |
| Chinese (Simplified), two lines | One extra character in 23 (4.3%) |
| Korean, two lines | 13.8% wrong, with two stray fragments. The weakest of the languages we tried |
| A page upside down, or turned a quarter either way | No wrong characters, once the page turned it upright |
| A page tilted 3 degrees, and one tilted 8 degrees | 0.4% wrong at 3 degrees. At 8 degrees most of the text was wrong |
| Fonts that look like handwriting | Easy ones (separate letters): 0 to 2% wrong. A flowing formal script font: 54% wrong |
We checked the page's own reading against a separate program built from the same engine, running on a computer, with the same language model. On our test page both gave the same result, with no wrong characters. We did not test every language in the list: Italian, Portuguese, Dutch, Polish, Turkish, Indonesian, Vietnamese, Ukrainian and Traditional Chinese use the same engine and the same kind of model, but we have not measured them.
Real photos are harder than these tests. A photo of a page taken on a phone has shadows, a curved sheet, uneven light and blur, and the results are worse. Plan on checking the text, and see the tips below.
Tips for better results
- Start from the best picture you have. The reader works best when letters are at least about 20 pixels high, and a sharp scan or a screenshot beats a photo. The page enlarges a small picture up to three times and shrinks very large photos (over 3200 pixels on the longest side) so that reading does not take minutes.
- Choose the right language, and the right picture type for a single line or scattered words.
- Cut away what is not text. A picture with only the text in it reads better. Use the crop tool first.
- Straighten a tilted picture. A few degrees are fine, but more than that is not. The rotate tool can straighten a page to the angle you choose.
- Fix poor contrast or color. The color fixer can lift a dull scan.
- Reading a PDF? This page takes pictures, not PDF files. Turn the pages into pictures with PDF to JPG, at 300 dpi for the best result, and add those here.
What it cannot do
- Handwriting. The reader is built for printed text. Neat, separate handwriting-style fonts worked in our tests, but joined-up writing did not, and we did not test real handwriting. Expect poor results.
- Layout. The result is text, not a formatted document. Columns are read in order, but tables, boxes and fonts are not rebuilt, and a table comes out as lines of words.
- Very small or blurry text, text on a busy background, and stylized logo lettering are often wrong or missed.
- A page tilted by much more than a few degrees. Turn upside-down and sideways pictures by hand if the automatic turn does not catch them, and straighten tilted ones.
- Understanding. It reads letters. It does not know what the words mean, so a wrong letter that makes another real word will not be flagged.
Languages
English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Turkish, Indonesian, Vietnamese, Russian, Ukrainian, Arabic, Hindi, Chinese (Simplified and Traditional), Japanese and Korean. Chinese and Japanese are returned without the spaces that the reader puts between their characters, and Korean keeps the spaces between words. Text written from top to bottom, and the many other languages that Tesseract supports, are not offered here.
Where the reading comes from
The reading is done by Tesseract, an open-source text recognition engine (Apache-2.0 license) that began at Hewlett-Packard, was released as open source in 2005, and is now maintained by a community of contributors, run in your browser through Tesseract.js (also Apache-2.0). We use its most accurate language models, in the compact form for the web. The engine, the worker and the language files are kept on this site, so nothing is loaded from another website.
Privacy
- The text is read inside this tab, and your files on disk are never changed or deleted. The text you get is shown in the page and saved only if you press Download.
- For what the site collects and keeps, see our privacy policy.
Frequently asked questions
How do I extract text from an image?
Add the picture to this page, choose the language of the text, and press Read the text. The text appears in a box under the picture, and you can copy it or download it as a .txt file. Nothing needs to be installed.
What is OCR?
OCR means optical character recognition: a program that looks at a picture of text and works out which letters it shows, so that the words become text you can copy, search and edit. This page does it with Tesseract, a well-known open-source engine.
Which languages can it read?
Nineteen: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Turkish, Indonesian, Vietnamese, Russian, Ukrainian, Arabic, Hindi, Chinese (Simplified and Traditional), Japanese and Korean. You can add a second language for a picture that mixes two.
Why is the text wrong, or full of strange characters?
The most common cause is the wrong language, so check that first. After that: the picture may be too small or blurry, tilted by more than a few degrees, or have a busy background, and fancy fonts and handwriting are hard. Crop to the text, straighten the picture, use a sharper or larger one, and try a different picture is choice. The confidence figure under each result tells you how sure the reader was.
Can it read handwriting?
Not reliably. It is built for printed text. In our tests, neat handwriting-style fonts with separate letters were read well, but joined-up script was mostly wrong, and we did not test real handwriting.
Does it keep the formatting, columns or tables?
No. You get plain text, with lines kept and a blank line between paragraphs. Columns are read in order, but tables, fonts and layout are not rebuilt. Switch on Join the lines of each paragraph if you want flowing paragraphs.
Can I read many pictures at once?
Yes. Add as many as you like, and press Read the text once. Each picture gets its own box, and Copy all the text or Download as .txt collects everything, with the name of each picture as a heading.
Can it read text from a PDF?
Not directly, because this page takes pictures. Use PDF to JPG to turn the pages into pictures at 300 dpi, and then add those pictures here.
Why was the first picture slow?
The first time, the page downloads the reading engine (about 4 MB) and the language data (about 1 to 3 MB). Your browser keeps the language data for next time, so a later visit starts faster, and reading a normal page then takes a second or two.
Is my picture sent anywhere to be read?
No. The reading happens in your browser tab, using files that this site provides. For what the site collects and keeps, see our privacy policy.