Image Redaction - Mosaic, Blur, and Black Bars

What does this image redaction tool do?

It hides the parts of an image that should not be shared. Drag a box over a card number, paint across a row of faces, or drop a black bar on an address, then export a flattened copy. Nothing is uploaded: the image is decoded, masked, and re-encoded inside this browser tab, and the export is a new bitmap rather than an overlay you could peel off later.

Key features

• Two ways to select: drag a rectangle, or paint freely with a round brush.
• Three ways to hide: mosaic with an adjustable block size, blur with an adjustable radius, or a solid bar in any color.
• Strength is stored as a percentage of the image's shorter edge, so a region hides just as much on a 1080p screenshot as on a 24-megapixel photo.
• Mosaic blocks align to one global grid, so neighbouring regions never look stitched together.
• Blur samples pixels from outside the region, so edges do not stay sharp.
• Region list with per-region effect, click to select, Delete to remove, and exact X/Y/W/H fields in percent.
• Move and resize any region, undo and redo with Cmd/Ctrl+Z, and hold the compare button to see the original underneath.
• Zoom to 2× or 4× for small text, with pointer accuracy unaffected by zoom.
• Auto-detect sensitive text with in-browser OCR, verified by checksum: bank cards (Luhn), Chinese resident IDs, IBANs (mod-97), mainland China mobile numbers, and email addresses.
• Auto-detect faces with the YuNet model, which runs locally through ONNX Runtime and is served from this site.
• Every detection is a proposal: review the list, uncheck what you do not want, then apply the ones you keep.
• Scan the open image or the whole queue in one run, with per-image review lists, a red badge on thumbnails that still need attention, and a single “redact all” step. A running batch can be stopped between images.
• Batch queue of up to 100 images, each with its own regions, exported individually or as a ZIP.
• Export as PNG, JPEG, or WebP; EXIF, GPS, and other metadata are dropped by the re-encode.

How to use it

1. Drop images onto the upload area or click to browse.
2. Choose the box or brush tool, then choose mosaic, blur, or a solid bar.
3. Drag or paint over everything sensitive. Adjust block size, blur radius, or brush size as you go.
4. Optionally scan for sensitive text or faces — on this image or across every image you added — then review what the models propose and apply the detections you agree with.
5. Switch to the select tool to fine-tune: drag a region, pull its handles, or type exact percentages.
6. Hold "compare" for a last look at the original, then download the redacted copy.

Which effect is actually safe?

Strongest first: a solid bar replaces the pixels outright and cannot be recovered. A large mosaic averages each block down to one color, so the detail is gone as long as the blocks are big — a few pixels per block is not enough, especially for digits. Blur is the weakest choice: a small radius keeps much of the original signal, and deconvolution can bring back readable text. For ID numbers, bank cards, license plates, and addresses, use a solid bar or a coarse mosaic. The editor flags regions whose setting is too light.

Does redaction remove hidden data too?

Partly, and in a useful way. The export is drawn on a fresh canvas and re-encoded, which drops EXIF, GPS coordinates, camera serial numbers, and any other metadata the original carried. What it cannot do is remove information that is still visible in the pixels, so check the whole frame: reflections, window titles, browser tabs, notification banners, and the thumbnail strip of another app often leak more than the part you meant to cover.

How reliable is the automatic detection?

It is a helper, not a guarantee. OCR reads what it can, and only strings that pass a real checksum are proposed, which keeps false positives rare: an order number or a price is not offered as a card number because it fails Luhn. Face detection works the same way — a proposal you confirm. The opposite case is the one to plan for: small type, low contrast, handwriting, rotated text, stylised fonts, profile views, and partly covered faces are missed regularly, and a miss is not redacted. Treat every detection as a first pass, then check the whole frame yourself.

Runs stay local. The OCR engine and its language data (about 7 MB for English and digits, 10 MB with Chinese) and the ONNX runtime used for faces (about 5 MB) are downloaded from a CDN the first time you use each feature and then served from the browser cache; the YuNet face model itself (227 KB, MIT licensed) is served from this site. Only those model files are fetched — the image never leaves the tab, and manual redaction works with no network at all.

Privacy and security

Decoding, masking, and encoding all happen locally through the Canvas API. No image is uploaded, nothing is stored after the page closes, and object URLs are released when you remove a file or leave. That matters more here than in most tools: the images people redact are exactly the ones they cannot afford to send to a server.

FAQ

Can I move a region after drawing it?

Yes. Switch to the select tool, click the region, then drag it to move or pull a handle to resize. Brush strokes move as a whole. Every gesture is one undo step, so Cmd/Ctrl+Z steps back cleanly.

Why does my mosaic look coarser on a small image?

Block size is a percentage of the shorter edge, which keeps the visual result consistent across sizes: 2% of a 1000px edge is a 20px block, and the same 2% on a 400px edge is 8px. The tool also enforces a minimum of 2px, because a 1px "block" would hide nothing.

What happens to the parts I did not cover?

They are copied through untouched at full resolution. Only the regions you drew are replaced, and only inside their own shape — a brush stroke affects the painted band, not its bounding box.

Why was a number on my screenshot not detected?

Detection needs OCR to read the characters first, and then the string has to pass its checksum. Anything blurry, tiny, rotated, or drawn in a decorative font tends to fail the first step; numbers that are not real cards or IDs fail the second by design. Draw a box over those by hand.

How does a batch scan work?

Set the scope to all images and the scan walks the queue one image at a time, reusing the engine it already loaded. Each image keeps its own review list, so nothing is applied behind your back: thumbnails with unreviewed detections get a red badge, and you can either review them image by image or use "redact all images" once you trust the results. "Stop" ends the run after the image currently being read, and everything found so far stays for review.

How do the automatic detections choose their effect?

They use whatever effect and strength the toolbar currently shows, so switch to a solid bar first if you want the strongest option applied to every accepted detection. Faces get generous padding around the detected box, and text regions get a smaller margin so neighbouring content stays readable.

Can I redact several screenshots in one go?

Add up to 100 images, click a thumbnail to switch between them, and redact each one. The badge on a thumbnail shows how many regions it already has, and "download all as ZIP" exports every image that has at least one region.