» About     » Archive     » Submit     » Authors     » Search     » Random     » Specials     » Statistics     » Forum     » Twitter     » RSS Feed     Updates Daily

No. 2901: Where's Garfield? Cross-Correlation Image Search

First | Previous | 2017-04-28 | Next | Latest

Where's Garfield? Cross-Correlation Image Search

First | Previous | 2017-04-28 | Next | Latest

Permanent URL: https://mezzacotta.net/garfield/?comic=2901

Strip by: Alien@System

{A Sunday Garfield strip run through a strange filter}

The author writes:

After my previous two entries dealt with auto-correlation and how it allows us to spot periodicity in pictures, here we can see how cross-correlations can be used to find objects in a picture. This picture was created by cross-correlating the Sunday strip with a mask, consisting of Garfield's head from panel five. Based on how cross-correlation works, we expect a bright spot every time our mask is above Garfield's head, since then the picture and the mask obviously fit well together.

And indeed, we see three bright spots, in panels two, five (no surprise) and six. We also see bright-ish splots in panels three and four, also centred on Garfield's face. However, there is no bright spot in panel one, six or the title panel. That's because cross-correlation, while useful, is of course not a "smart" algorithm. It just shifts one picture over another, without rotating, flipping or re-sizing them to get a match.

To find out how that is done, you might want to talk to Google. In their attempt to make a useful picture search (where you could give it, say, a picture of a corgi and it would search the web for other pictures of corgis), they had to deal with these advanced kinds of image processing. Their result was a neural net approach, which means nobody knows from the outside how exactly the program determines what's a dog and what's a pipe (or a picture of one), but I bet that at some deep level, cross-correlations are still involved. As their Deep Dream visualisation of the "seeing" process of their neural net shows, they have managed to create a program which can't just identify a dog pretty well, but also see this dog in a cloud.

If you're sick of cross-correlations now, that's okay. I'm done, but slightly proud that it's been seven years since SRoMG #130, and we've still not run out of maths things to play around with.

[[Original strip: 1981-09-20.]]

Original strip: 1981-09-20.