Photo2PatternFree license plate blur
Blur a license plate — with a floor it will not go under
Box each plate in the photo and cover it with a solid rectangle, a Gaussian blur or a grid of blocks. The difference between this and every other blur tool is that the weak settings are not available: the blur will not go softer than the plate’s own length divided by 12, the block grid will not go finer than three blocks along it, and choosing something under either floor disables the download and tells you which number is wrong. It is free, asks for no account, and the photograph is opened in the browser. A plate is seven high-contrast characters in a known typeface, which is the one thing you must never attenuate when you could delete.
- 100% free
- No signup
- Weak settings refused
- Solid fill by default
- Plate shown at 4×
Press Ctrl+V anywhere on the page to drop in a listing photo you copied from a tab — no need to click the field first.
How to cover a plate so it stays covered
If the fill you choose cannot hold at the size you publish at, the tool stops you before the export rather than after.
Box every plate in the frame, not only your own
Drag a rectangle over each registration in the shot — front and rear if both are visible, the parked car behind yours, the one reflected in a store window, the trade plate on the transporter. Take in the whole plate including the border and the fixing bolts rather than stopping at the characters, because a covering that leaves the first digit poking out has covered six characters of seven and the seventh is worth more than the six.
Take the solid rectangle unless you have a reason not to
It is the default because it is the only one of the three with a proof behind it: the pixels are replaced by one value and there is nothing left to work back from. The blur and the block grid attenuate rather than delete, which means they hand over a degraded copy of the characters, and a degraded copy of a seven-character string is a puzzle with an answer. Choose one of them when the photo needs to look natural, and read the numbers the page puts under it.
Read the enlargement, then export
Each plate is redrawn underneath at up to four times its size with smoothing off, so you are judging the actual exported pixels rather than a thumbnail that flatters them. Beside it sit the deviation applied, the width of one character at this plate's size, and how much of the character signal is still getting through. If a setting is below the floor the page states which of the two numbers is wrong, offers the corrected value in one click, and refuses to produce a file until it is taken.
Technical specifications
| Blur floor | Deviation of at least the plate's long side over 12; anything softer is refused and cannot be exported |
|---|---|
| Pixel ceiling | 3 blocks along the plate, because each block hands over one average |
| Where the floor comes from | A character pitch p of the plate's length over 7 — the count a US and a UK registration both settle on — with exp(-2·π²·s²/p²) held below 1/255, giving s > length/13.2 and rounded to length/12; then checked against the real kernel, which agrees, since 12 measures 1 part in 602 and 13 measures 1 in 254 |
| Solid fill | #111827 across the whole box; no residue, and the enlargement shows a flat rectangle |
| Blur kernel | Five box passes per axis, clamped to the box itself rather than the surrounding photo, so the output is a function of the plate alone; three passes bottom out at 1 part in 217 and never clear the floor |
| Proof view | Each plate redrawn at up to 4× with smoothing off, alongside its deviation and pitch |
| Working resolution | 2,000 pixels on the long side, where a plate runs 100 to 300 pixels |
| Metadata in the export | None — GPS, camera serial and capture time do not survive the re-encode |
Frequently asked questions
Is a light blur enough to hide a plate?
No, and the arithmetic says how much is. A Gaussian of deviation s reduces a repeating pattern of period p to exp(-2 x pi squared x s squared / p squared) of its strength. Characters on a plate repeat at roughly a seventh of the plate's length, so the period you have to kill is length over 7 — and pushing that below the one part in 255 an 8-bit channel can still hold needs a deviation above 0.53 of the period, which works out at the plate's length over 13.2. This page rounds that to length over 12 and will not export anything softer. A deviation of a twentieth of the plate, which looks thoroughly blurred on screen, still leaves about a tenth of the character signal standing — twenty-five times the level an 8-bit channel stops resolving, and a residue that size is exactly what sharpening exists to find.
Why does it refuse a setting my photo editor is happy with?
Because a photo editor is applying a filter and this page is making a claim. An editor has no idea what is under the brush, so it cannot tell you whether a 6-pixel blur has destroyed the content or merely softened it — it will apply the same 6 pixels to a plate 80 pixels long and one 400 pixels long, and only one of those is covered. The setting here is dimensionless for that reason: it is the plate's own length divided by a number, so every plate in the frame gets a covering scaled to its own size, and the floor is a statement about the result rather than about the brush.
Can a pixelated plate really be read back?
Yes, and it has been demonstrated publicly — Bishop Fox published a working reconstruction of pixelated text in 2022, and a similar tool against pixelated screenshots appeared in 2020. The method is simple enough to explain: pixelation is deterministic, so an attacker who knows the typeface renders candidate strings, pixelates them on the same grid, and keeps the ones whose block averages match. It works because each block hands over one number and a short string does not need many numbers to be pinned down. A plate carries seven characters drawn from about 36 symbols, which is roughly 78 billion possibilities — a large space, but not against twelve or twenty averages. This page caps the grid at three blocks along the plate, where one block spans more than two characters, and even that is a mitigation rather than a proof.
Does covering the plate make the car anonymous?
It removes the identifier that has a public register behind it, which is the important one and is not the only one. A car is described by its make, model, trim, color, wheels, roof box, window stickers, a dent in the wing and a dealer surround, and any of those can place a specific vehicle for somebody who already knows the neighborhood. The photograph carries more still: a house number, a street sign, a delivery label on a parcel in the trunk, a reflection of the person holding the camera in the paintwork. Cover the plate first because it is the one anyone can look up, then look at the rest of the frame.
Does the file still say where the photo was taken?
Not once it comes out of here. Phones write GPS coordinates into the photograph alongside the camera make, the serial number and the timestamp, and those tags outlive most edits — which makes covering a plate and posting the original file a strange trade, since the plate needed a lookup and the coordinates did not. Everything you download from this page is encoded from raw pixels, so none of that metadata is carried across. It is a consequence of building the export from a canvas rather than an option that can be left switched off.
Do I have to cover other people's plates as well?
If you are publishing the photograph, treat every legible plate in it the same way you treat your own. A registration is a key into a public record, and the person whose car happened to be parked behind yours did not choose to be in the shot. This is a settled expectation in the used-car and dashcam communities and the reason the page takes a list of boxes rather than one — the common case for a driveway photograph is two plates, and for a street scene it is four or five.
What about a dashcam clip rather than a still?
This page covers one frame and cannot help with a video, and the distinction matters more than it sounds. A minute of dashcam footage at 30 frames a second is 1,800 stills, the plate is legible in a good many of them, and covering it in the one you exported changes nothing about the file you are actually posting. Video redaction needs the box tracked across frames and the whole thing re-encoded, which is a different tool. Export a still if a still is what you want to share.
About why a plate is the worst thing to blur lightly
Everything that makes a registration plate easy to read at fifty yards makes it easy to recover from a bad covering. The characters are high contrast by regulation, they are drawn in a typeface fixed by the issuing authority, they sit at a known pitch on a known background, and there are seven of them. That last number is the one that matters. Seven characters drawn from roughly 36 symbols is about 78 billion possible strings, which sounds enormous until you notice that an attacker does not have to search it blind: they can render every candidate in the right font, degrade it exactly the way you did, and keep the ones that match. Blur and pixelation are both deterministic, so this works, and it has been done — Bishop Fox published a working reconstruction of pixelated text in 2022, and a tool against pixelated screenshots circulated in 2020. What defeats it is not a stronger filter but a different operation. A solid rectangle replaces the pixels with a constant; there is no degraded copy for anything to be compared against, and no amount of computation recovers a string from a value that never depended on it.
Where a soft covering is unavoidable, the strength can at least be derived instead of guessed, and both numbers on this page are. Take the blur first. A Gaussian of standard deviation s attenuates a spatial period p by exp(-2·π²·s²/p²) — a Gaussian’s transform is another Gaussian, which is never zero, so a blur is in principle invertible and what actually stops the recovery is the point at which the surviving signal drops below the quantization step of the file. Character pitch on a seven-character plate is its length over 7. Setting the attenuation at that period below one part in 255, which is all an 8-bit channel can represent, needs s above 0.53 p, or the plate’s length over 13.2. Rounded the safe way, that is length over 12, and it is what the slider will not go under. That derivation describes a true Gaussian, though, and what runs is a cascade of box passes approximating one — so the floor is checked a second time against the cascade’s own transfer function, per plate, and a setting that fails that check is refused whatever the slider says. Measuring it changed the build twice. Three box passes, the usual recommendation, never reach one part in 255 at any setting, because a box cascade has sidelobes where a Gaussian has none; five do. And with five, the weakest setting that holds for every plate length from 20 to 1,200 pixels is exactly one part in 12 — one part in 13 measures 1 in 254 and fails — so the derived floor and the measured floor turn out to be the same number. Now the pixel grid. Pixelation leaves exactly one average per block, so the count of blocks across the plate is the count of numbers an attacker gets to test candidates against; at seven blocks a block sits inside a single character and the reconstruction is straightforward, so this page stops at three, where one block spans more than two characters. Three is a mitigation and the interface says so. Only the solid fill is a proof.
The last part of the job is not arithmetic. A plate is the identifier with a public register behind it, so it goes first — but a photograph of a car in front of a house carries a house number, a street sign, a permit in the windshield, a VIN plate visible through the glass, a delivery label, and the photographer reflected in the door. Every legible plate in the frame counts too, including the neighbor’s and the one in a store window, which is why this page takes a list rather than a single box. When the things you are covering are people rather than numbers, the face page proposes the ovals and asks you to check them. When several different kinds of thing have to go from the same picture — an address, a signature, a barcode and a face — the redaction page holds them as one list and is the page that explains why the covering has to be flattened into the exported pixels. And when nobody is hunting for what is underneath and you only want the look, pixelating the frame hands you the block size with no floor attached, in the same way cutting an image down to a few colors changes a picture without concealing anything in it.
Product and company names shown here are the trademarks of their respective owners. Photo2Pattern is an independent tool and is not affiliated with, sponsored by or endorsed by any of them.
Where the photo of your car is opened
Your picture is opened by JavaScript running in the tab you are reading this in. It is never uploaded, never sent to a server and never stored anywhere — there is no upload step to undo, because there is no upload. Close the tab and nothing of it remains.
Worth spelling out on this page in particular: a photograph you were about to hand to a website in order to have its plate hidden would have arrived at that website with the plate still in it. Nothing here is transmitted, so that trade is not being made.