How to use Geulppopgi, with measured accuracy

It reads text out of an image or a screenshot and gives it back as plain text. It reads Korean and English. Handwriting is not supported.

The reader is Tesseract, open source, running from files this site hosts itself.

Large images are shrunk so the longest side is 3000 px. Small images are not enlarged. Files up to 20 MB are accepted.

Press Pick an image, drag an image onto the page, or paste a copied image with Ctrl+V. Taking a screenshot and pasting it is the fastest way.

Reading starts as soon as the image is in. When it finishes, the text appears in a box you can edit.

Text to read and reading mode

If Korean and English are mixed, choose Korean and English. For an image with only English, English only can fit better.

Auto is the default. For a photo of one page or one block of text, One block of text sometimes helps. Switching re-reads the same image.

Using the result

Copy puts the text in the box on your clipboard. Save as .txt writes the same text to a UTF-8 file.

Make a card builds one 1080x1350 PNG. It holds the character count, never the text itself.

These numbers come from three sample images we made, read by this app exactly as you use it. One image per type, so treat them as a rough feel. They are not a guarantee.

Accuracy is how many characters match the known answer after removing spaces.

Results

Printed document, Auto: 98.1%. One synthetic page of clean text on white.

Messenger screenshot, Auto: 81.3%. One synthetic chat screen with colored bubbles.

Photo of paper, One block of text: 79.2%. One synthetic image with tilt, shadow, blur and noise added. It is not a real camera shot.

Same photo read with Auto: 23.4%. On this sample, switching to one block of text helped a lot.

How to read these numbers

Real photos vary in light and focus, so expect lower. Compare important text against the original after reading.

If a re-measurement with the same method gives different numbers, we fix this list.

Text size and contrast decide most of the result. Check these in order.

Straight on, well lit

Shoot paper straight on, not at an angle. Keep shadows off the letters.

Move closer for small text

If letters look tiny, step closer and shoot again. Do not zoom screenshots, drop in the original.

Only the part you need

If a cover or a picture is mixed in, crop to the text. One block of text mode can fit better.

Your image is processed in this browser only. It is not uploaded, and the text it yields is not sent anywhere.

The reader and the Korean and English language files are files of this site. Nothing comes from an outside server. Korean plus English downloads about 10 MB the first time and English alone about 8 MB, then the browser keeps them.

The server only counts step names, such as the page was opened, an image was added, or reading finished.

We use Tesseract.js, Tesseract.js Core and the tessdata_fast language files. All are under the Apache License 2.0.

Full license texts are at /ocr/LICENSE-tesseract.js.txt, /ocr/LICENSE-tesseract.js-core.txt and /ocr/LICENSE-tessdata-fast.txt on this site.