How does facial detection with matrices work?

[Superposed_Cat]

Hi all, there is a facial detection program called eigenfaces that supposedly uses eigenvectors to recognise faces, can anyone here share any intuition on how that works or send a link? Any help apreciated.

 

[BruceW]

http://en.wikipedia.org/wiki/Eigenface
http://en.wikipedia.org/wiki/Principal_component_analysis

The rough idea is that you start with a bunch of images of faces, and you compare the images, to calculate 'eigenfaces', which are composite faces which vary from the mean in different ways. The critical point is that these composite faces contain as much information per image as possible. So, say you are given 100 images of faces, then you could build new faces using a mix of all 100 faces. But, instead, you could calculate the eigenfaces, and just use the first 5 eigenfaces. And by using a mix of these 5 eigenfaces, you could potentially make (almost) as much variety as you could with the original 100 images. So, in a sense, we are trying to reduce the dimensionality of our problem, while still keeping as much of the information as possible.