Image to Text

Optical character recognition looks at the shapes in a picture and decides which letters they are. It does that very well when the letters are large, dark, sharp and in straight lines, which is why a screenshot or a flatbed scan of a printed letter usually comes out nearly perfect. It does less well as soon as the picture adds problems of its own: a page photographed at an angle, a shadow across one corner, a crease, a faded receipt, or text so small that each letter is only a few pixels tall.

Resolution is the most common cause of trouble. The documentation of the open-source Tesseract engine says recognition works best on images of at least 300 DPI, and the same rule of thumb applies to OCR in general. A phone photo taken close to the page usually has enough pixels; a small image saved from a chat app or a website often does not. Scaling a small picture up does not add the missing detail, so when you can, go back to the original or scan again.

Some sources are easier than others. Screenshots of websites, apps and slides are the best case, because the text was drawn by a computer and has perfectly sharp edges. Flatbed scans of office documents come next. Phone photos of books are harder, since the page curves toward the spine and the lines bend with it; pressing the book flat and shooting from straight above helps a lot. Receipts, labels and signs are the hardest common case, with thermal print that fades, shiny surfaces and text in several sizes on the same piece of paper.

The list below sums up what to aim for when you photograph or scan a page for text.

Printed text follows rules: each letter of a font always looks the same, so the recognizer has a stable pattern to match. Handwriting has no such rules. The same person writes an "a" differently from one word to the next, letters run into each other, and lines drift up and down the page. Setting Text type to Handwriting switches to a recognizer trained for that variety, and it handles neat block letters and clearly separated words fairly well.

Be realistic about the rest. Joined-up cursive, doctor's notes, faint pencil, writing on lined or patterned paper, and notes crammed into margins are hard for any OCR system, and parts may come out wrong or not at all. Expect to correct the result by hand, and keep the original image next to it while you do. For typed forms that were filled in by hand, the printed labels will usually read well while the handwritten answers need the most checking.

Mixed pages are common: a printed form with handwritten answers, or a typed letter with a signature and a few notes added by hand. Choose the text type that covers most of what you need. Printed mode is more accurate on typed text, so if only the signature is handwritten, leave it on Printed.

Plain text is the quickest way to get the words into an email, a chat or a form, and it opens everywhere. Word keeps paragraphs, columns and, with Keep pictures on, the photos and logos from the page, so it is the format to choose when you want to edit and reuse the content. Searchable PDF leaves the scan exactly as it was and adds selectable, searchable text behind it, which is ideal for archiving paper records or making a scanned contract searchable. Markdown gives clean headings and lists for notes apps and wikis.

Language choice also shapes the result in ways that are easy to miss. A page in French read as English loses its accents, and a page in Hindi or Arabic read with the wrong language gives little usable text at all. Right-to-left scripts such as Arabic and Hebrew depend on the matching language to come out in reading order. If a page mixes scripts, for example English product names inside a Hindi letter, set the main language first and add the second one under Also recognize.

PDF input can have many pages. Under Pages, read every page, only the first, or a range such as 2-4, which saves time when the text you need is on one page of a long scan. If you have a document that is already digital, such as a PDF made from Word, the text is usually stored in it already and the Document Converter can pull it out without recognition. For more ways to work with documents and scans, see all the document tools.

Recognized text can include personal details such as names, addresses and account numbers, so treat the output with the same care as the original paper. Files are processed for the conversion only and deleted afterwards; the How it works page describes the steps.