Pyramid coder with nonlinear prediction

Antoine MANENS (manensa9@leland.stanford.edu)
Laurent MEUNIER (lmeunier@leland.stanford.edu) 

The starting point will be the Burt-Adelson pyramid coder. This is a 
two-steps algorithm : first, a low-pass filter is applied to remove
pixel-to-pixel correlation. Then, the difference (which has low
entropy) is coded. 
Hence, there are two areas of improvements
1- the filter kernels can be optimized so that the error image has lower
variance. The starting point will be the Gaussian filter used by
Burt/Adelson.
2- the error images have some kind of relation and some further
optimization can be performed : I'll look in particular at nonlinear
prediction.