Yes I'm using a similar method in the first section of the article (information theoretic solution), but I think that reddit is using the entropy of an image instead of the self-information. The self-information yields to better results in our case (empirically). In the latter sections I use different methods.

This looks like a nice, simple solution. Now I wonder if there is a solution similar to this written in PHP.

I found this:

http://jueseph.com/2010/06/opticrop-content-aware-cropping-w...

Seems to be a similar solution using PHP and ImageMagick.