How does OCR turn a picture into text?
Image to text conversion, known as optical character recognition (OCR), is software that looks at a picture of text and writes out the characters in it. This page does it in your browser, and the image is not uploaded. A photo of a receipt is only a grid of colored dots to a computer; OCR is the step that decides which of those dots are the letters "T", "o", "t", "a", "l".
This page uses Tesseract, an open-source OCR engine that began at Hewlett-Packard in the 1980s and 1990s and has been developed in the open since 2005. Version 4 added a neural network that reads whole lines of text at a time instead of single letters, and that is the engine used here. It reaches your browser through tesseract.js, a version compiled to WebAssembly so it can run inside a web page.
First the engine converts the picture to black and white so letters stand out from the page. Tesseract's own documentation says this step can go wrong when the background is unevenly lit, which is why shadows cause so many bad results. Next it finds blocks of text, splits them into lines and cuts each line into words. Each line is then handed to the neural network, which outputs the most likely characters along with a score for how sure it is. That score is the confidence number you see on this page, from 0 to 100 for each word.
What affects OCR accuracy?
The engine is the same every time, so what changes the result is the picture. These are the factors that move accuracy most, with what to do about each one.
| Factor | What goes wrong | What helps |
|---|---|---|
| Resolution | Letters only a few pixels tall blur into each other | Tesseract recommends about 300 DPI; zoom in or rescan |
| Contrast | Gray text on a gray background gets lost in black-and-white conversion | Dark text on a light background |
| Lighting | Shadows and glare make part of the page darker or washed out | Even daylight, no flash straight on glossy paper |
| Tilt | Lines that slope make word and line detection fail | Rotate so the lines are level |
| Cropping | Page edges, hands and desk edges are read as characters | Crop to the text with a thin margin |
| Language | Wrong language data gives wrong letters, such as accents lost | Pick the language the text is written in |
| Fonts | Script, decorative and handwritten styles are far less reliable | Printed text in a plain typeface |
Source: Tesseract documentation, Improving the quality of the output (checked November 9, 2026). The right-hand column is our own summary of it.
Screenshots are the easy case: the text is sharp, perfectly horizontal and high in contrast. Photos of signs, book pages and whiteboards are harder because of perspective, curvature and light. Handwriting is a separate problem altogether; Tesseract is built for printed text, so treat any handwriting result as a rough guess.
How do you get a cleaner result?
- Crop first. Cut the picture down to the text you want with our image cropper. Dark scanner edges and background objects are the commonest source of stray characters.
- Check the language. A French menu read as English still gives words, but accents and many letters come out wrong. The picker is above the image.
- Straighten the page. If lines slope, rotate the picture in any photo app before adding it. Text that is a few degrees off often reads fine; 15 degrees usually does not.
- Use the biggest version you have. The page enlarges images smaller than about 1,000 pixels on the long side, but enlarging cannot add detail that was never captured. A screenshot at full size beats a thumbnail, and if the file is huge, our image compressor can bring it under the 30 MB limit.
- Fix light text on dark. Tesseract's documentation advises dark text on a light background for version 4 and later. If your image is inverted, invert it first.
- Read the doubtful boxes. Dotted red outlines mark the words the engine was least sure of. Check those against the picture and leave the green ones alone.
What do the outlines and confidence mean?
Each word the engine finds is drawn on your image as a box. The box style comes from that word's confidence score. The cut-offs below are our own reading guide, not an official scale.
| Outline | Confidence | How to treat it |
|---|---|---|
| Solid green | 85 to 100 | Usually correct; a quick look is enough |
| Dashed orange | 60 to 84 | Check against the picture, especially names and numbers |
| Dotted red | Below 60 | Often wrong; retype from the image |
Source: Cut-offs are our own guide. Scores come from Tesseract's per-word confidence, 0 to 100.
A high score does not guarantee the word is right. The engine is confident about what it saw, not about what you meant, so a clean scan of a typo stays a typo. Numbers deserve a second look whatever their score: an 8 and a 3, or a 1 and an l, are the pairs that most often swap.
What is downloaded and what is not?
The picture you add is never uploaded. It is decoded by your browser and handed to the OCR engine running in a web
worker on your own device. What does cross the network is the engine itself. The first time you read an image,
your browser downloads three things. For English they come from this site (the /ocr/ folder). For
any other language, the language data comes from the public jsDelivr content delivery network,
cdn.jsdelivr.net, which is tesseract.js version 7's default source:
| File | Approx. size | Kept for next time |
|---|---|---|
| Worker script (tesseract.js) | 0.1 MB | By the browser cache, subject to its normal rules |
| Engine core (WebAssembly) | About 4 MB | By the browser cache, subject to its normal rules |
| Language data (.traineddata) | 0.7 to 3 MB each | In your browser storage (IndexedDB), so a language downloads once |
Source: tesseract.js 7.0.0 source and documentation; sizes checked on November 9, 2026.
Like any request, a download from the CDN shows jsDelivr your IP address and which files you asked for, and so which language you chose; with English, nothing leaves this site. No download includes the image or any text read from it. After the first read in a language, the language data comes from your browser's storage; the engine files may too, depending on your browser's cache. Clearing your site data removes the stored language files. Apart from those downloads, the site only counts page views.
The image and the recognized text stay in this tab's memory until you close or reload it. Nothing is saved on our side, because there is no server step to save to. The site does count page views, as every page here does. Close the tab when you are done with an ID or a bank statement.
Which image types can it read?
| Type | Works here | Notes |
|---|---|---|
| PNG | Yes | Best for screenshots; transparency is drawn on white |
| JPG / JPEG | Yes | Photos and scans; heavy compression blurs small text |
| WebP | Yes | In any browser that can display it |
| GIF, BMP | Yes | First frame of an animated GIF |
| HEIC / HEIF | Sometimes | Only where the browser can open it, such as Safari; otherwise convert it first |
| No | Export a page as an image first |
Source: Behavior depends on your browser's own image decoder. tesseract.js itself reads bmp, jpg, png, pbm, webp and non-animated gif.
The page first draws the image onto a canvas, so anything your browser can display is accepted, and the phone camera's rotation flag is respected. For iPhone HEIC photos that your browser cannot open, use the HEIC to JPG converter. PDF pages can be turned into images first with our PDF to JPG converter. Images up to 30 MB are accepted.
Common questions
Is the image uploaded anywhere?
No. The image is read on your device. Only the OCR engine and the language file are downloaded, from this site for English and from the jsDelivr CDN for other languages, and only the first time.
Why does the first read take longer?
The engine and the language data, a few megabytes together, have to arrive before anything can be read. Later reads in the same language skip the language download.
Can it read handwriting?
Not reliably. Tesseract is trained on printed text. Neat block capitals sometimes work; cursive almost never does.
Why are the accents or characters wrong?
The language is probably set to English. Choose the language the text is written in and the engine will read again with the right letter set.
Can I read two languages in one image?
This page reads one language at a time. Tesseract itself can combine languages, but that needs more data and tends to lower accuracy, so for mixed text run the image once per language and keep the better lines.
What does the confidence percentage mean?
It is the average of the engine's own score for each word, weighted by word length. It says how sure the engine is, not what share of the text is correct, so use it as a prompt to proofread rather than as a grade.
Sources
- Tesseract documentation: Improving the quality of the output
- Tesseract documentation: Data files (best, fast and standard models)
- Tesseract OCR project on GitHub
- Smith, R., An Overview of the Tesseract OCR Engine (Google Research)
- tesseract.js project on GitHub
- tesseract.js API documentation (default paths, caching, output formats)
- tesseract.js documentation: Local installation (hosting the files yourself)
Last reviewed November 9, 2026.