Auto Redact (Text)

Auto Redact finds the text in a photo, shows you a box around each piece it found, and covers the boxes you keep with a solid box, a blur or pixelation. It runs on your device, and your image is never uploaded.

It finds text, it never reads it

The model is a text detector. Given an image, it returns a map of where text is likely to be, and that map is turned into boxes. There is no recognition step anywhere in this tool: nothing turns those pixels into letters, so no word, number or name is ever extracted, stored or sent anywhere.

That is a property of the model, not a promise about how the output is handled. A detector has no way to say what the text was.

What it downloads

The first time you press Find text, the page downloads one model file: the PP-OCRv3 text detector, 2,423,490 bytes (about 2.3 MB), published by OpenCV under the Apache 2.0 licence. It is fetched from a fixed, pinned version, checked against a fixed fingerprint before it is used, and cached by your browser so later runs start without the download.

The detector runs on ONNX Runtime, the same engine the site's other model tools use. If no earlier tool in this browser has loaded that engine, and you have not used the offline pre-download, the first run also fetches it from this site: up to about 24 MB, once.

According to its authors' technical report, the detector was trained on about 127,000 images: real scene photos from public text-detection datasets and web image search, plus synthetic images of long, angled and tabular text.

The steps

1. Choose an image. A very large photo is first scaled down to the working limit this device can hold safely, and the page tells you when that happens.
2. Press Find text. Standard looks at the image at up to 960 pixels on its long side. Fine looks at up to 1536, which helps with small text in a large photo and takes longer. Neither enlarges a small image.
3. Review every box. Red boxes will be covered. Click a box, or untick it in the list, to leave it alone.
4. Choose a treatment and press Cover.
5. Read across the result at full size, then download the PNG.

Which treatment removes the text

• Solid box replaces the pixels with a flat dark rectangle. Nothing of the lettering remains in the file. It is the default, and the one to use for anything that matters.
• Blur mixes each box's pixels together. The amount is set by the height of the letters in that box, so large lettering gets a proportionally larger blur than a fixed blur would give it.
• Pixelate replaces each box with blocks about as tall as its letters.

Blur and pixelation keep an average of the original pixels. They stop a casual reader, but published techniques have recovered pixelated text from screenshots, so do not treat either as removal.

What it gets wrong

We checked the detector on nine images from our own test set before building this page: a dashboard screenshot, a shop front with a painted sign, a photographed invoice, a contact card, a low-contrast label, a bus with a destination board, a landscape with no text, a hillside with sheep, and a pile of crumpled receipts. It boxed every line on the screenshot, the card and the label, the shop sign, and the bus board and its number plate. It found nothing in the empty landscape. It also got three things wrong, and they are the three to watch for:

• It can box things that are not text. A row of distant trees on the hillside was boxed with high confidence. Switch those boxes off before you apply.
• It can miss text. The crumpled receipts produced no boxes at all: print a few pixels tall on creased paper was below what it sees. On the invoice it boxed the main printed lines but missed several small, faint marks near the top of the angled page. Try Fine, and if text is still missed, cover it by hand in Privacy Blur.
• Slanted text gets oversized boxes. The boxes are upright rectangles, so on a page photographed at an angle each box also covers some of the page around the line. For redaction that is the safe direction to be wrong in.

It covers text and nothing else

Faces, licence plates, reflections, the photo's metadata and simple context can all still identify a person or a place. Privacy Blur has faces, plates and this same Text option in one review, plus boxes you draw yourself. The Metadata Remover strips location and camera details from the file.