Image to Text Converter (OCR)

Copy the text out of a photo, scan or screenshot. Auto-enhance fixes small text, dark mode, shadows and sideways pages before reading, and the whole job runs in your browser: the image never leaves your device.

Converter Media & Files Updated Oct 14, 2026
How to Use
  1. Drop a photo, scan or screenshot on the drop box, click it to browse, or paste a screenshot with Ctrl+V. Or press one of the samples under Presets.
  2. Leave Enhancement on Auto and press Recognize. The first read downloads the OCR engine and the English model (about 8 MB, then cached by your browser); later reads start straight away.
  3. Check the picture: every word gets a box coloured by confidence, green at 85% or more, amber at 60–84% and red below 60%. Red boxes are the words to proofread.
  4. Fix anything the reader got wrong directly in the text box, then press Copy text or Download .txt.
  5. Open Show Work to see each pass auto-enhance tried (enlarged, inverted, shadows removed, rotated) and why the winning read won.
Image
Presets: sample images
Extracted Text
Your image appears here, then every word the reader finds gets a box: green for confident, amber for unsure, red for words to check.
Confidence
—
Words
—
Characters
—
Time
—

Worked Example

These are the presets above, read by this page’s engine with the results copied out unedited. Accuracy is 1 minus the character edits needed to turn the output into the original text, divided by the original’s length, with line breaks and runs of spaces counted as one space.

A dark-mode error screenshot. Five lines of 12 px light-grey text on a near-black background, 560 × 104 pixels. Read as it is, Tesseract gets 96.4% (7 wrong characters out of 194), turning [email protected] into react8ls.3.1 and [email protected] into react-dong17.0.2. Auto-enhance finds the median word box in that first read is only 8 px tall, so it enlarges the image 4× (40 ÷ 8 = 5, capped at 4) and inverts it, because the border is dark. That read scores 90% mean confidence and 98.5% accuracy; the 3 slips left are 1 read as l twice and ^ read as ~.

A receipt. A clean 400 × 420 scan in a monospaced font. The first, unenhanced read already scores 91% mean confidence and a score of 133.6. The best enhanced read scores 134.9, less than the 3% it would need to replace the original (0.97 × 134.9 = 130.9), so the original read is kept. Every price, the Tax 8% 0.89 line and the TOTAL 12.04 come back exactly. The one error in 175 characters is the masked card number, **** read as ***x: 99.4%.

The common mistake: trusting the confidence on a shadowed photo. A 900 × 500 phone photo of a notice, tilted 2.5°, with a hand shadow darkening its left side. With enhancement off, Tesseract reports a reassuring 71% confidence, but the text it returns is fragments such as LE, ICE and li close early at 4:30 pm. Only 45.3% of the characters are right, because the words in the shadow were dropped rather than misread. Tesseract’s automatic-layout read is even more misleading: 93% confidence on just 11 words. With Auto, the background is divided out and the image enlarged 2×; the winning read (“background removed + denoised”) finds all 33 words at 96% confidence, a score of 143.7 against the original’s 62.5, and gets all 181 characters right, including 4:30 pm and Friday 24 October.

Show Work

Load an image and press Recognize (or press a preset) to see every pass the reader tried and why the winner won.

How Auto-Enhance Works

Text size
f = 40 ÷ median word height
Word heights come from the first read. f is kept between 1× and 4×, and enlarging never takes the image past 16 megapixels.
Dark mode
border grey < 110 → invert
Light text on a dark background is flipped to dark on light, which is what the LSTM engine expects.
Shadow removal
I′ = 255 × I ÷ B
B is the local background: the brightest pixel within about 0.9 line heights, then blurred. Shadows and tinted paper divide out to white.
Black and white
Otsu: max w0w1(μ0 − μ1)²
The threshold that best separates ink from paper. “Ink only” thresholds a second time inside the dark pixels, so grey stripes and watermarks drop out.
Picking the winner
score = Σ conf × characters
Over words read at 50% confidence or more. The original read stays unless another scores at least 3% higher.
When to stop
mean ≥ 88% stop · < 70% rotate
After two enhanced reads, the rest are skipped once the best read is 88% confident. Below 70%, the page is also tried at 90°, 180° and 270°.

From HP Labs to a Browser Tab

The engine behind this page is Tesseract. It was developed at Hewlett-Packard Laboratories in Bristol, UK, and at HP in Greeley, Colorado, between 1985 and 1994, and was entered in the University of Nevada, Las Vegas (UNLV) 1995 Annual Test of OCR Accuracy. Development then stopped until HP released the code as open source in 2005. Google developed it from 2006 until August 2017; Ray Smith was the lead developer until 2017, and Stefan Weil leads it today.

Tesseract 4.0.0, released on 29 October 2018, added a neural-network engine built on LSTM (long short-term memory) layers that recognises a whole line at a time instead of matching one character shape at a time. Version 5.0.0 followed on 30 November 2021. Tesseract can read more than 100 languages, each from its own trained model; this page uses the English model from the “best” set, which works only with the LSTM engine.

Tesseract.js wraps a WebAssembly build of the same engine so it runs inside a web page. This page uses tesseract.js 5.1.1 with a Tesseract 5.1 core, served from this site. The engine’s own documentation explains why the image matters as much as the engine: accuracy drops below 10 pt text at 300 dpi (a lowercase x about 20 pixels tall), and the LSTM engine wants dark text on a light background. Those two fixes, size and polarity, are the first things auto-enhance applies.

Sources: Tesseract README (history, versions, languages); Tesseract 4.0.0 release; R. Smith, “An Overview of the Tesseract OCR Engine”, ICDAR 2007 (UNLV test, 2005 release); Tesseract docs, Improving the quality of the output and the old FAQ on text size; tesseract.js README.

About This Tool

This converter turns a picture of text into text you can copy, edit and search: a photographed page, a scanned letter, a receipt, a screenshot of an error, a sign. It opens 323 image file extensions through the site’s image engine, including HEIC from iPhones, TIFF scans, PSD and camera RAW, not only the handful a browser can show.

What sets it apart from running Tesseract once is auto-enhance. It reads the image up to about a dozen ways (as it is, with automatic layout, enlarged, inverted, with shadows divided out, cleaned to black and white, rotated) and keeps the read with the most confident text, then puts two-column pages back in reading order. The word boxes and Show Work let you see exactly what it read and where it was unsure.

Everything runs in your browser. The image is decoded and read on your device by a WebAssembly build of Tesseract served from this site; neither the picture nor the text is ever uploaded. The first read downloads about 8 MB of engine and English model, which your browser then keeps.

It suits anyone who needs the words out of a picture: office work with receipts and forms, students with photographed handouts, developers with screenshots of errors, and readers who want text they can enlarge or have read aloud.

Related tools: PDF to Image, Image Cropper, and Universal Image Converter.

Frequently Asked Questions

Is my image uploaded to a server?

No. The image is decoded and read inside your browser tab by a WebAssembly build of the Tesseract OCR engine, which this site serves from its own folder along with the English model. Nothing about the picture or the text is sent anywhere, so you can use it on receipts, letters and ID documents.

How accurate is it?

On clean print it is close to perfect: the sample receipt came back with 1 wrong character out of 175 (99.4%), and a 12 px dark-mode terminal screenshot with 3 wrong out of 194 (98.5%). On a set of 26 deliberately hard images (tiny text, shadows, noise, stripes, sideways and upside-down pages) auto-enhance raised accuracy from 76% to 94%. Handwriting, heavy blur and photos where the letters are only a few pixels tall are where it still struggles.

Why can the confidence be high when the text is wrong?

Tesseract reports confidence for the words it found, not for the words it missed. Read with enhancement off, the shadowed sample photo scored 71% mean confidence while getting only 45% of the characters right: the shadowed half of the page was simply dropped. That is why auto-enhance ranks reads by confidence × characters, and why a short result from a full page is worth a second look.

Which languages and file types does it read?

English, using Tesseract’s LSTM “best” model (the language menu lists only the models actually installed). Images are decoded by the same engine as the site’s image converter, so it opens 323 file extensions: JPG, PNG, WebP, GIF, HEIC, AVIF, TIFF, PSD, SVG, camera RAW and many older formats. It does not open PDF files; turn the pages into PNGs with PDF to Image first.

How do I get the best result from a photo?

Shoot the page straight on, in even light, close enough that the text fills the frame, and crop away anything that is not text. Tesseract’s own guidance is that accuracy drops below 10 pt text at 300 dpi, which is a lowercase x about 20 pixels tall, and that below 10 pixels there is very little chance. Auto-enhance enlarges small text up to 4×, which is how the sharp 12 px screenshot sample still read at 98.5%, but it cannot add detail the camera never captured: a blurred photo of small print stays blurred.

Common Use Cases

Error messages in screenshots

Paste a screenshot of a terminal or an error dialog and copy the message into a search. A 12 px dark-mode npm error read at 98.5% after auto-enhance enlarged it 4× and inverted it, against 96.4% read as-is.

Receipts and expenses

Pull the items, prices and totals off a receipt for a spreadsheet. On the sample receipt every price, the 8% tax line and the 12.04 total came back exactly; the only slip was one * of a masked card number.

Photos of notices and handouts

A page photographed with a hand shadow across its left side read 100% with auto-enhance and 45% without, because the shadow is divided out before the text is read.

Two-column articles

Newsletters and textbooks come back left column first, then the right, instead of every line running across both. In the two-column sample all 320 characters came back in order; the only error was one stray ‘ mark before the first heading.

Scanned PDFs

Render each page to PNG with PDF to Image at 4× (288 dpi, close to the 300 dpi Tesseract’s documentation recommends), then read the pages here one at a time.

Reading aloud

Extract the text from a photographed letter or a poster and paste it into Text to Speech to have it read out, at whatever speed suits you.

Last updated: