Image to Text

How to copy text from a photo

1. Choose one or more photos or screenshots (JPG, PNG, WebP; HEIC where your browser opens it), drop them on the box, paste one with Ctrl+V or Paste a picture, or on a phone Take a photo.
2. If the photo shows a page on a table, the page is found by itself and outlined. Tap Adjust the page to move its corners, or choose Whole picture.
3. Leave the language at Telugu + English (mixed lines are read too), or choose one.
4. Tap Read the text. Each picture shows its progress; Stop keeps what has been read.
5. Choose Lines as in the picture or Join into paragraphs, correct anything in the text box, then Copy, Share, or save it as a .txt or Word file. Several pictures come out as one text, in order, with each picture’s name before its part (in Word, each from a new page).

ఫోటో ఎంచుకోండి, “Read the text” నొక్కండి, వచ్చిన తెలుగు అక్షరాలను కాపీ చేసుకోండి: అంతా మీ ఫోన్‌లోనే.

What it does before reading

• Flattens a page photographed at an angle, from its four corners, as a scanner app does.
• Turns it the right way up: a page photographed sideways or upside down is found and read upright.
• Straightens a slightly tilted page (up to 15°).
• Takes out shadows and uneven light, so a shadow across the page or a yellowed newspaper reads like a clean scan; light text on a dark screen (dark mode) is read too.
• Sizes the letters for the reader: a 12-megapixel phone photo is made smaller, small text in a screenshot larger, so letters are about the height the reader is trained on.
• Removes specks and the lines of boxes and tables, which OCR otherwise reads as “I”, “1” or “|”.
• Keeps the reading order: two newspaper columns come out one after the other, not mixed line by line; a table’s cells are kept apart by tabs.
• Checks the Telugu: the text is proper Unicode (NFC, no stray zero-width joiners).
• Fixes common misreads with Telugu Typing’s word list and letter model (a vowel sign of the wrong length, a vattu missed, look-alike letters), only where the reader was unsure and one word clearly fits; names, numbers, English and words in the list are never changed. Words it is still unsure of get suggestions.
• Warns before reading when a photo is blurred, its letters are too small, it is too dark, washed out by glare or turned too far, so you can take it again (or read it anyway).

Proofread shows each line read under the strip of the photo it came from, the changed words underlined and the doubtful ones marked. Tap a word for suggestions, or type it in English letters (roju → రోజు, as in Telugu Typing); Tab goes to the next doubtful word.

On our test set of 24 pictures (news, a G.O.-style notice, a newspaper page in two columns, screenshots in light and dark mode, and photos of paper at an angle, sideways and upside down), these steps took the mistakes from about 19 in every 100 Telugu letters to under 1, and from 7 in 100 English letters to none. Most of the gain is on photos taken at an angle or sideways and on the two-column page; clean pictures read about as well either way.

Better Telugu reading (a switch, on by default on computers) has a second Telugu reader look again at the lines Tesseract was unsure of, and keeps the better word of the two, using Telugu Typing’s word list. On our hand-checked pages it took the Telugu words read wrong or missed from 57 in 749 to 25; it makes reading up to three times as long and needs up to about 250 MB more memory, so it starts off on phones. It downloads about 22 MB once, from ap7am, and is kept with the models; if it cannot run (an older browser, too little memory), the text is read with Tesseract alone and the page says so.

Everything happens in your browser: your pictures are never uploaded to ap7am or anywhere else. The first time, your browser downloads the reader and its Telugu and English models from ap7am, about 6 MB, and keeps them (the same ones OCR PDF uses); “Forget the downloaded models” lets them go, Better Telugu reading’s too.

What it cannot do

• OCR makes mistakes, more often in Telugu: letters with a vottu (్య, ్వ, ్ర), long and short vowels (ి / ీ), and small, blurred or faint print. Check names, numbers and dates against the picture before you use them in a form or a message.
• Handwriting is not read, nor fancy lettering in posters and wedding cards.
• A page bent like an open book, or a photo taken very close at a steep angle, reads worse: hold the phone flat above the page, in good light.
• PDFs are not read here: for a scanned PDF use OCR PDF; for a PDF whose Telugu copies as wrong letters, Telugu PDF to Unicode.
• Reading takes a second or two a picture on a computer, and a few seconds on a phone.

Questions

Made with Tesseract OCR (tesseract.js and tesseract.js-core, Apache 2.0 licence) and its Telugu and English models (Apache 2.0, packaged under the MIT licence), and for Better Telugu reading harsha-desaraju’s Telugu OCR model (MIT licence) run by ONNX Runtime Web (MIT licence), all running in your browser. Licences.