Get the words out of a photo, a scan or a PDF and edit them
Everything here runs in your browser. Files are never uploaded.Text recognition reads the letters in a picture and gives you characters you can search, copy and correct. It is what you need when the words you want are inside a photograph of a page, a screenshot, or a scan that your PDF reader will not let you select. The recognition runs inside your browser, so a photographed marksheet, contract or prescription is never uploaded.
A photo or a screenshot, and you want the words. Image to Text gives you plain text to copy.
A photographed or scanned page, and you want to edit it. Image to Word gives you a .docx with the recognised text in it, which is the right choice when you are going to correct and reuse the document rather than quote from it.
A PDF that will not let you select text. That PDF is a scan: the pages are pictures. Export the pages as images and run Image to Text on them.
A PDF that does let you select text. Do not use recognition at all. PDF to Text pulls out the characters that are already in the file, perfectly and instantly, and PDF to Word keeps the structure as well. Recognition would only introduce mistakes into text that is already exact.
Clean printed English at a sensible resolution is read with very few errors. Accuracy falls with everything else: a photograph taken at an angle, a shadow across the page, a fax-quality scan, a stamp over the text, small print, handwriting. Handwriting is not supported in any useful sense, and no consumer tool reads it reliably.
The single biggest lever is the picture you feed it. Shoot the page square-on in even light with the whole page in frame, or scan at 300 dots per inch. A better photograph beats every setting.
You get the words. You do not reliably get the layout: columns, tables, headers and the position of a signature are reconstructed approximately or not at all. Always read the output against the original before you use it for anything that matters, because a recognition error is silent. A misread digit in an account number looks exactly as confident as a correct one.
No. The recognition engine is downloaded to your browser the first time you use it and then runs on your device. A photographed marksheet, payslip or ID never leaves your machine.
English is the default and the most accurate. Other scripts, including Hindi, are supported by loading that language model, and accuracy in every language depends far more on the quality of the picture than on the script.
No. These engines are trained on printed type. Handwriting comes back as nonsense, and that is true of most free tools rather than a limitation of this one.
Because a scan is a photograph of a page, so there are no characters in the file to search, only pixels arranged to look like letters. Recognition is what turns those pixels into characters.
Scan at 300 dots per inch, or photograph the page square-on in even light with no shadow. Crop to the text, keep the page flat, and avoid enlarging a small blurred picture, which adds no detail.
Always. Recognition errors are silent and confident: a 5 read as an S looks no different from a correct character, so check anything you are going to rely on, especially numbers.