Image to Text Converter

What it reads well, and what it does not

The engine behind this page is Tesseract, the long-established open-source recognition engine, compiled to WebAssembly so it can run inside a browser. It is very good at one thing — clean printed text — and honest results depend on knowing where that stops.

Pick the language the text is written in, even for a short snippet. Each language has its own model, and the English one does not know accented letters. Reading the French line “À côté de l'hôtel” with the English model gave A c6té de I'hétel in our test; with the French model it came back exactly right.

The confidence badge on each result is the engine's own average certainty for the words it found. Treat anything under about 60 per cent as a draft to check line by line against the picture, and anything over 85 as normally clean.

Stray marks. Icons, logos and borders in a screenshot are sometimes read as letters. Words the engine itself scores as near-certain misreads, and lines made up only of such marks, are left out of the result. Everything it is reasonably sure of stays in — including its mistakes — so still check names and numbers.

Your image stays on your device

Many image-to-text websites send the picture to their server, read it there and send the words back. That is the easy way to build one, and it means a copy of your screenshot — a bank statement, a contract, a private conversation — has passed through somebody else's machine.

This page works differently. When you add the first image, your browser downloads the recognition engine and a model for the chosen language — about 5 MB with the English model, less for the others — and from then on everything happens in the tab. The picture is decoded by your browser, prepared on a canvas, and read by the engine running on your own processor. There is no server-side part to this tool, so there is nothing that could receive the image even by accident.

Two practical consequences. The first image takes a few seconds longer while the engine arrives; after that the browser keeps it, so the next image — and your next visit — starts straight away. And there is no daily limit, no account and no queue, because nothing you do here costs a server anything.

Getting a clean result

Tesseract's own guide to improving output quality comes down to a handful of points. This page handles five of them for you and leaves the rest to the picture you give it.

• Resolution. The engine works best on images of at least 300 dpi, so small text needs more pixels. The page measures how tall the letters in your image are and enlarges it, up to three times, until typical letters are about 22 pixels high. On a full-HD screenshot of this very page, that alone turned sigmawirenet into sigmawire.net and Download txt into Download .txt.
• Dark text on a light background. Since version 4, Tesseract expects it. A dark-mode screenshot is inverted as a whole. White text on a coloured button, badge or banner is handled on its own: each solid dark shape with lettering inside is turned into dark text on white, and the rest of the image is left alone. In the same screenshot, the white-on-pink Choose images button had come out as : oose images before this step. If the detection guesses wrong on a mixed image, set Background yourself.
• Even lighting. A photo lit from one side has a grey half and a white half, and a single dark-or-light cut-off loses the text in the grey half. The brightness of the paper is estimated across the image and evened out first. On a test page shaded from dark grey to light grey, the engine returned nothing at all before this step and every line after it.
• A margin round the text. Text that touches the edge of the image is read badly, so a 10-pixel white border is added before recognition. You can crop tightly without worrying about it.
• Straight lines. The engine needs lines of text to run horizontally. A page tilted by up to 15 degrees is measured and straightened before reading. It matters more than you might expect: in our tests, three lines of text tilted by 5.5, 8, 10 and 12 degrees each came back empty from the engine as they were, and word-perfect once straightened. Upside-down pages are not detected, and sideways ones not always: the Rotate button on each result turns the image a quarter at a time and reads it again. Perspective, where a photo taken at an angle makes the lines converge, is not corrected, so hold the camera square to the page.
• Crop out what is not text. Large photos, logos and scanner shadows around the edge get mistaken for letters. A tight crop reads faster and cleaner.

Which layout to choose

Before it reads anything, the engine divides the image into blocks, lines and words. The Layout setting tells it what kind of image to expect. Each option maps to one of Tesseract's page segmentation modes.

If a result has whole lines missing, switch layout before anything else; changing it re-reads every image already on the page.

Line breaks and paragraphs

The text comes out with the line breaks of the original, which is right for code, addresses and lists and wrong for prose you want to paste into a document. Join lines into paragraphs merges the lines inside each paragraph and keeps the blank lines between them. A line that ends in a hyphen is joined without a space and the hyphen is kept, so “well-known” survives intact; a word that was only split for the line break will keep its hyphen and is worth a glance.

What this converter cannot do

• PDF files. It reads images only. For a PDF use PDF to Text OCR, which takes the text layer straight out of digital PDFs and only runs recognition on scanned pages.
• Handwriting, beyond neat capitals.
• Formatting. Output is plain text. Tables come out as words separated by spaces, and bold, italics and font sizes are lost.
• Columns side by side. Two columns of text can come out interleaved, the first line of the left column followed by the first line of the right. Crop each column and read it as its own image.
• Spaced-out capitals. Menu-style text in widely spaced capitals can lose the gap between words, so ALL TOOLS comes back as ALLTOOLS. That happens inside the engine and enlarging the image does not change it.
• Other languages. Six are offered: English, Spanish, French, German, Portuguese and Italian. Text in another script, such as Arabic, Chinese or Hindi, will not be recognised with these models.
• Maths and symbols beyond ordinary punctuation.
• Formats your browser cannot open. The image is decoded by the browser, so HEIC or TIFF files work only where the browser itself can display them. Converting to JPG or PNG first always works.

Need to share the picture itself rather than the words in it? The Image to URL Converter gives an image a direct link you can paste anywhere.

Questions

Related tools

Browse all Sigma Wire tools - every free tool on the site, grouped by category.