Hacker News Reader

Comment by erikb | Hacker News Reader

In some cases even blurring faces might be a bad idea. Just because we are unable to unblur a face today doesn't mean we are unable in 10 years or 100 years. In many cases this might not be a problem but in some cases this might lead to trouble later on which can be avoided just as easily.

View on HN

You cannot unblur mosaic blurring. It's the equivalent of a hash. The best you can do is brute force possible input vectors. There will be many collisions. This technique only works because digits/numbers limit the input space for a credit card or bank number. For faces, the best you can do is validate if someone you already suspect or someone you have in a database, is the origin of the mosaic. If you had a picture of every person in your country, you could run them all through and find the origin - but you're not really unblurring as much as extracting information to do process of elimination on a 'image hash'

I can imagine you can computationally unblur mosaic on video, if the object (e.g. face) is not changing, just the camera is moving around a little.

You can in fact do this, it is a technique known as Super-Resolution and has been around for at least a decade. http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=5...

The most precise maps of Pluto were made with a similar technique: http://www.boulder.swri.edu/~buie/pluto/mapstory.html

That's really cool. I wonder if that could be used to get images of other planets.

gus_massa posted a link about Pluto:

http://www.boulder.swri.edu/~buie/pluto/mapstory.html

Maybe the scientifically correct definition of "to unblur" is reverting the process directly by applying a mathematical algorithm. In a normal conversation like what we have here, bruteforcing a good enough result can also be considered "to unblur" because for the person it's the same result: Everybody knows who it is.

The un prefix means to reverse. If you have no database of people you're not going to be able to do anything in the way of reversing the mosaic. It's not a reversal as much as it is a heuristic brute force.

You can't use the brute force method to find any "person" you have in the database. You can only find any "photograph" that you have in the database.

If you have a photograph of me, and I just let my hair and beard grow for a month, then you won't match the mosaic version of my hairy version with the old photograph.

Perhaps you can add some filters to adapt the contrast, luminosity, hue, illumination, rotation, pixel shifts and noise. It would be more difficult to fix the head orientation, open/close mouth/eyes.

Imagine a future where 3D depth-sensitive cameras are prevalent - this means that a 3D model of a person's head can be easily obtained.

Then futuristic analysis software that, from a picture, can determine all the light-sources in that image and 3D positions thereof.

So take a photo, analyze it for light-sources, then brute-force it by applying 3D models of every face in your database with the known light-sources and run it through the mosaic filter.

Futuristic technology could help here. Probably wouldn't work since I'm sure one person's mosaic would match too closely somebody with similar features and a similarly-shaped head.

i agree with erikb.. In theory at least, it seems possible to me that unbluring mosaic could happen if you have in multiple mosaic images from different angles (like mosaic in a video) even without a database with the possible faces).

Even with that, the amount of information is reduced from 1000s of pixels to less than 100. It's possible to extract some information but not enough to do a reversal

If you have trial faces at the right angle, you can unblur faces, it's just a bit harder to generate. But nothing major is going to change in 10 or 100 years; it's just information theory as to whether you can identify the person/number or not.

Indeed. See http://en.wikipedia.org/wiki/Christopher_Paul_Neil

Swirl is not a lossy process. If he had blurred his face he would have been much harder to catch.

It's partially lossy, from the look of it (you can see circular artifacts), but you're right that enough is recoverable to make it meaningless in terms of hiding data.

Reply on news.ycombinator.com