🚀 go-pugleaf

I've been thinking about pattern recognition with fractals too, but I
always come to a basic problem with the block-wise fractal transformations:

The coefficients are not robust with
respect to perturbations of the image. Two images that are very similar
can have very different affine coefficients. The reason, if you haven't
seen this yourself, is that in the search for domain blocks the algorithm tries 
to find the domain block with the smallest error. Two domain blocks from wildly 
different parts of the image can have close and small, errors. Consequently, a 
small change in the image can favor one domain over  another. The affine 
coefficients of these two domains can be very different. So a small change in 
the image can result in a large change in the affine coefficients.

I conclude that the block-wise fractal transform is not a very good candidate
for pattern recognition applications.  Any comments?

Is anyone aware of other fractal transforms that might be better? 

Fernando Pineda (fernando@aplcomm.jhuapl.edu)