Image to Text Converter

This image to text converter reads the printed or typed words in a picture — a screenshot, a photo of a document, a scanned page — and gives them back as plain text you can copy, edit or save. It is optical character recognition (OCR) that runs entirely in your browser, so it suits anything you would rather not upload: payslips, contracts, ID documents, private chats.

Recognition is done by Tesseract, the open-source OCR engine, in its tesseract.js 7 build compiled to WebAssembly. It uses Tesseract's LSTM engine — a neural network that recognises a whole line of text at a time rather than one letter at a time — with the English language model. The engine, its worker and the model are served from this site, not a third-party CDN, and run in a background worker in your tab. The first run downloads about 7 MB (the engine is about 4 MB and the English model about 3 MB); the model is then kept in your browser's storage, so later runs start faster. Only English is available at the moment.

Before recognition the image is drawn onto a white background, so a screenshot with a transparent background does not turn into black text on black. If the image's longer side is under 1,000 pixels it is enlarged 2×: Tesseract's own documentation says it works best at 300 DPI or more, and text in a normal-size screenshot is often only 12–16 pixels tall, which is at the edge of what it reads reliably. Very large images are reduced to fit within 25 megapixels and 16,384 pixels per side — so a very long full-page screenshot is shrunk to 16,384 pixels tall, and the result says so; if its text becomes too small to read, crop it into parts.

OCR is an estimate, not a copy. Accuracy depends mainly on resolution, contrast and font: clean, dark text on a plain light background at a decent size comes out close to perfect, while small, blurry, low-contrast, stylised or skewed text produces mistakes such as 0 read as O, l as 1, or missing punctuation. Tesseract is designed for printed text, and handwriting usually comes out badly. Layout is not reproduced — you get the words in reading order, not the columns, tables, fonts or colours. The confidence figure shown with the result is Tesseract's own average certainty over the words it found, not a measured accuracy, so proofread names, numbers and amounts.

The "Join lines into paragraphs" option helps when you paste the result into a document. OCR keeps the line breaks of the picture, so a paragraph arrives as many short lines. With the option on, single line breaks become spaces, a word hyphenated across a line break ("infor-" / "mation") is joined back together, and blank lines between paragraphs are kept.

The image never leaves your device: it is decoded, prepared and recognised inside your browser. You can confirm in the developer tools' Network tab that the only requests go to this site for the engine files — the page also loads Google Analytics, but no request carries your image or its text.

1. The file's first bytes are checked to confirm it is really PNG, JPG, WebP, BMP or GIF (whatever its name says); files over 20 MB are refused. 2. The browser decodes the image (the first frame, for animated images) and draws it on a white canvas. If its longer side is under 1,000 px it is drawn at 2× size; if it exceeds 25 megapixels or 16,384 px on either side it is reduced to fit, with a note in the result. 3. A tesseract.js 7 worker is started with the English model and the LSTM-only engine (OEM 1). The worker script, WebAssembly core and gzipped model are all loaded from this site's /vendor/tesseract/ folder. 4. Tesseract binarises the image internally (Otsu's method), finds the text lines and recognises each line with its LSTM network. Progress is shown on the button. 5. The text is tidied: line endings normalised, trailing spaces removed, three or more consecutive line breaks collapsed to one blank line. With "Join lines into paragraphs" on, line-end hyphens before a lowercase letter are removed, other single line breaks become spaces, and blank lines stay as paragraph breaks. 6. The mean word confidence Tesseract reports is rounded to a whole percentage and shown with the word and line counts. The worker is shut down when recognition finishes.