This week’s edition of Media Watch, “Pixelating protects identity? Think again“, examines the threat image search engines pose to anonymity. Drop a disguised photo into Google images and the chances are you will find the original in the search results.
Intrigued, I thought I would try it out. The pixellated the photo of Tom Waits was my second test. The first image I found to try was a golden pyramid. (It is from a presentation I recently pulled together on cognitive dissonance, but that is unlikely to be a helpful explanation).
In this case the search results came close to being artistic: an impressive array of alternative golden pyramids.
I can see that Google images could be rather fun.
Possibly Related Posts (automatically generated):
- Dissonance and Debt (1 September 2011)
- Cognitive dissonance (26 June 2011)
- Following one link too few…a mea culpa (28 May 2010)
- Pyramid Perversion – More Junk Charts (12 March 2010)
Well, I’ll be damned…
Some friends I have are used to make fun of my “paranoia”. I’ll see to it they watch this segment, just so I can ask them who’s laughing now?
Laughing magpie?
And squawking, too!
It picked up the strangely edible texture of your pyramid, which is even more impressive than picking up the missing bits of Tom’s face.
@Golden Orb: shades of Toblerone? Mind you, not sure how edible the scrubbing brushes, bottom right are!
It would be interesting to know what there algorithms do. A reasonable engineering starting point would be to perform fourier transforms on the full image and selected sub images. Then restructure the output so that it is in order of major of major to minor features, so that comparisons are fast, as they stop if major features are not in common. Then all you need is to determine a measure of how close the images are. Presumably all the pixelated images are matching well on sub images, and giving some commonality to the overall image. These days there are lots of other options like fractals and hidden markov models to do what is effectively dimension reduction.
Would it then be able to match, say, a pixelated single face taken from a group shot? Or only if the overall pictures are the same?
It seems to have quite a high tolerance for error, which could positively match the wrong person. Although it is somewhat heartening that it puts up several options if it is unsure.
@Golden Orb: I haven’t experimented with group shots, but I suspect it may fail. What I have tried was submitting a photo of my own face. It was a photo which is not online anywhere and the suggestions Google came up with were rather amusing!
One possibility is that they look at sub images, so as to allow for cropping.
It appears it gets the colors and shades farily accurately. However, geometric shapes show much more variance… Interesting experiment.