A horse of a different color

In my Color Shift experiment, I rotated all three colors: red to blue, green to red, blue to green. It is also possible to hold one color constant while switching the other two. There are two ways to shuffle three colors, and three ways to shuffle two colors, for five ways in all. Alternatively, there are six permutations of three objects; the five just described plus the original colors equal six.

With computers, when I come to a fork in the road like this, my inclination is to try all possibilities and see what happens. It doesn't take much more programming, and CPU time is free.

The picture above is the most interesting of the five. I explain this by noting that the background (grass, shrubs) is mostly green, and the foreground (horse) is mostly red, so holding the green constant and switching the red and blue results in a blue horse on a normal-looking background. The dissonance between the foreground and the background is what makes the picture interesting.

This is all very glib, but of course the computer doesn't know which is the foreground and which is the background. There are plenty of optical illusions to show that this is not always so easy for humans to figure out either. I tried averaging all the pixels and found that the picture as a whole is slightly more red than green. Holding red constant and switching green and blue results in a normal-looking horse against a bluish background... which is not interesting. The other three transformations change both the foreground and the background, which is even less interesting.

Of course, what I am looking for is some kind of metric that will tell me which of the five color transformations will likely result in interesting pictures.