Image to Text Converter (OCR)

Optical character recognition (OCR) turns the letters in a picture back into text you can copy, search and edit — a receipt, a slide, a screenshot of an error message, a page from a book. This tool reads images in English, Korean, Japanese, Chinese and eight other languages directly in your browser with the open-source Tesseract engine, so the pictures stay on your device.

How to use this tool

1. Choose one or more images. Straight, well-lit, high-contrast pictures give the best results.
2. Pick the language of the text; for mixed Korean and English pages choose “English + Korean”.
3. Leave the layout on “A single column” for screenshots and documents; choose Automatic for pages with several columns, or a single line, one block or scattered text when parts are missed.
4. For faint scans tick “black and white”; for white text on a dark screen tick “light text on a dark background”.
5. Press the button, then copy the extracted text. The first run downloads the engine and language data.

Formula

Mean confidence = average of the engine’s per-word confidence (0–100%) for the image

Before recognition the image can be enlarged 2× (small text below about 20 px tall is hard to read) and converted to pure black and white with Otsu’s method, which picks the grey level that best separates the two groups of pixels (it maximises the between-class variance). For Chinese and Japanese, the spaces Tesseract inserts between characters are removed.

Example

A phone screenshot of a two-line English error message, 1170 px wide, read with English and the default single-column layout, usually returns both lines word for word with a high mean confidence. The same text photographed at an angle under a lamp scores lower and may need a few corrections — a confidence below about 70% is the cue to check every line.

Limits worth knowing

Printed text works far better than handwriting, which is mostly not recognised. Tables come out as lines of text without their grid. Very small, blurred, curved or low-contrast text produces errors, so crop to the text and photograph it straight on. The engine (a few megabytes) and each language’s data (about 1.5–3 MB) download the first time you use them and are cached after that.