OCR - Image to Text

What OCR does

OCR (optical character recognition) reads the letters in a picture and turns them into text you can copy, edit and search. A photo of a page, a scanned receipt, a screenshot or a picture of a sign is just pixels to a computer. OCR finds the shapes of letters in those pixels and outputs the characters they represent.

This tool runs OCR in your browser with a WebAssembly version of the Tesseract engine and supports 12 languages. Your image is processed on your own device and is not uploaded; only the language data, which the engine needs, is downloaded the first time you use a language.

What determines accuracy

OCR accuracy depends far more on the picture than on the engine. A clear, well-lit, high-resolution image of printed text in a common font can reach very high accuracy, while a blurry photo taken at an angle can produce nonsense. The main factors are:

• Resolution. Text should be at least about 30 pixels tall for reliable reading; for a printed page, a 300 dpi scan is the standard target.
• Contrast. Dark text on a light plain background works best. Coloured or patterned backgrounds and low-contrast text cause errors.
• Straightness and focus. Photograph the page from directly above, in even light, and keep it sharp. Skewed and curved pages are harder.
• Font and content. Common printed fonts do well. Decorative fonts, handwriting, very small print and mixed languages do worse.

Getting better results

Preparing the image gives the biggest improvement. Crop away everything except the text with the Image Cropper, so the engine does not try to read borders and pictures. If the picture is small, enlarge it slightly with the Image Resizer; if it is huge, reduce it to a sensible size. Rotate it so the lines are level. Choose the correct language before you run the recognition, because the engine uses language data to decide between similar shapes, such as an "l" and a "1".

When the first result has errors, try again with a cleaner picture instead of correcting many words by hand. Then proofread the text yourself: OCR often confuses similar characters (0 and O, 1 and l, rn and m), and errors in numbers are easy to miss. The article how OCR works and how to get better results explains these mistakes in more detail.

Scanned PDFs and photos of documents

For a scanned PDF, first turn its pages into pictures with PDF to JPG, at 200 to 300 dpi for good OCR, and then run each picture through the tool. If you have photos of a paper document, combine them into a PDF with JPG to PDF afterwards. To turn recognised text into an editable file, paste it into a document; for PDFs that already contain real text, the PDF to Word converter is the better route.

Take care with sensitive documents. Running the recognition in your browser keeps IDs, contracts and medical papers off other people's servers, which is an advantage over many online OCR services.

Photographing pages with a phone

A phone camera can give good OCR results if you take a few precautions. Lay the page flat and shoot straight from above so the lines are parallel to the edges of the frame. Use daylight or even light and avoid shadows from your hand and glare from glossy paper. Tap to focus, move close enough that the text fills the frame, and take the picture at the highest resolution your phone offers.

Document-scanner apps flatten curved pages and boost contrast automatically, which often helps more than any setting in the OCR tool. If you only have a normal photo, cropping and slightly increasing the contrast before recognition can rescue a poor result.

Handwriting and other limits

Tesseract is designed for printed text. Handwriting recognition is much less reliable and the output will often need heavy correction. Tables and multi-column layouts are read as text lines, so their structure may be lost. Very long images or huge photos can be slow on older devices, and the first run for a language takes a little longer while its data downloads. Verify names, numbers and dates in the output before relying on them.