Google Begins Blurring Faces In Street View

mytrip notes a News.com article reporting that Google has begun blurring faces in its Street View service, which has spawned privacy concerns since its introduction last year. Google has been working for a couple of years to advance the state of the art of face recognition. Quoting News.com: 'The technology uses a computer algorithm to scour Google's image database for faces, then blurs them, said John Hanke, director of Google Earth and Google Maps, in an interview at the Where 2.0 conference...' Google wrote about the program in their Lat/Long blog."

default

[Anonymous Coward]

If you have an out of focus picture, can you manipulate the image mathematically to put it "in focus" or is there some information lost in the out-of-focusness so you can't do this.

A:Yes

And if so, with the appropriate app, will you be able to un-blur the people's faces in Google Street View?

A:Yes

Re:Can you focus out-of-focus pictures

[Lobster Quadrille]

You can't add pixels that aren't there, and an out of focus picture is effectively a lower resolution.

You can, however, apply statistical analysis and AI learning techniques to guess the likely locations of pixels. In that way, you can sharpen a photo somewhat, though it may be inexact. My understanding is that contextual analysis is the next step- if you have pictures of a person and a blurry person, and have more pictures of that person and less-blurry people, you can make predictions about who the fuzzy people are.

Of course, I wear a beard so that I'll always be fuzzy.

Re:Other uses for this technology

[taybay]

Google has already added a face search feature to image search. It's not too shabby either. I'm sure they're looking into other options as well.

Let me get the point...

[Tangamandapiano]

Aside from time factor (I suppose it works 24h/day), what's the big legal difference from what the TV programs do when they show random people, in scenes from the cities or so?

Google has been developing this for some time.

[lhaeh]

This article [blogoscoped.com] from a year ago shows that Google has had public implementations of facial recognition for some time. Simply appending &imgtype=face to a Google image search URL will just show images of faces.

Re:What privacy concerns?

[Fozzyuw]

That's an overly simplified view. Are you saying that in public it should be legal to be able to take pictures of anybody from any angle/viewpoint? (eg: upskirt)

Interesting that you should say that [bbc.co.uk]... as this was a recent BBC article I read. And it's not even "upskirt", it's just taking pictures of peoples behinds. Of course, the best part is the last sentence...

He might have some explaining to do when he finally gets home.

Re:Can you focus out-of-focus pictures

[pavon]

You can't add pixels that aren't there, and an out of focus picture is effectively a lower resolution.

No, it isn't. Think about an unfocused camera - all the light is still hitting film/CCD, it is just spread out. So from an information theory point of view you haven't lost any data, you just put it into another form. If you consider what would be a single point of light, the energy in that point is spread out in a normal distribution (aka bell curve, aka gaussian). So the blurred image is just all these Gaussians functions overlayed on top of each other. Computer blurring algorithms do pretty much the same thing.

From a signal processing perspective, this is the same as convolving with a Gaussian. And if you take the Fourier transform of that blurred image, you get the transform of the image multiplied by the transform of the Gaussian (which is just another Gaussian). From there all you have to do is divide by this Gaussian, take the inverse transform, and walla, you have the desired non-blurred image. This is called a deconvolution [wikipedia.org], and I've written code to do this for an image processing class.

There are some caveats. You have to guess how blurred the image is - what focal length is and what not. Noise and compression can kill you, so you need to filter those out first (or limit your deconvolution filter to low frequency content). In addition at the edges of the image (or edge of the blur boundary) information is genuinely lost as the gaussian falls outside the boundary and is discarded.

Focus Magic [focusmagic.com] is a commercial package that refocuses blurred images, and they have some interesting sample photos.

Re:Can you focus out-of-focus pictures

[Solandri]

You can't add pixels that aren't there, and an out of focus picture is effectively a lower resolution.

You can, however, apply statistical analysis and AI learning techniques to guess the likely locations of pixels. In that way, you can sharpen a photo somewhat, though it may be inexact. My understanding is that contextual analysis is the next step- if you have pictures of a person and a blurry person, and have more pictures of that person and less-blurry people, you can make predictions about who the fuzzy people are.

This is wrong. An out of focus picture is not lower resolution. All the original information is still there, it's just been smeared in a mathematically consistent manner - something called the point spread function of the lens [wikipedia.org] at that degree of misfocus. It's very possible to mathematically focus a misfocused picture [focusmagic.com] after it's been shot. The main barriers are not knowing the particular lens' exact point spread function, sensor noise (the de-convolution spreads the sensor noise to adjacent pixels), and grid resolution. But the site I linked to shows you can still get pretty decent results using a generic PSF.

Re:Kudos to Google!

[flink]

Well, I don't know... the one about blanking out maps of China sure seems to improve privacy.

Yeah, and try finding your way around Israel [google.com] using gmaps as well.

Well, that's just the thing

[Moraelin]

Well, the thing is: people are more than happy to jump to conclusions, without having any context for that photo.

E.g., I've waited for a taxi at a street corner before. Admittedly, I'm a guy, but I don't remember any law or moral code that forbids women to use taxis either. So it doesn't take too much of a stretch of imagination to allow for the possibility that those two girls too were just waiting for their ride. Or maybe they went shopping and are waiting for the BF of one of them to come give them a ride home. Or various other possibilities.

We don't actually have enough data to make a judgment there. If they're on the same corner for several hours straight, daily, yes, then they're probably working there. But we don't know that. We have just a snapshot that doesn't really say anything by itself.

But people are more than happy to jump to a conclusion anyway.

The same applies to a lot of other situations.

E.g., it's trivial to take someone's photo that looks like he's walking towards a brothel, when he's just really walking past it.

E.g., the most heinous case of "it's not what it looks" involved a UK chav filming himself pissing on what looked to him like a dead-drunk woman passed out on the side-walk. Turns out that she wasn't drunk, she was just dying of liver failure. (And before you jump to conclusions again, there _are_ ways to get that without being an alcoholic.) So instead of calling an ambulance, the retard filmed himself pissing on her while she was dying, and posted the movie.

Re:Kudos to Google!

[nbritton]

A better view [google.com]