Pyramid coder with nonlinear prediction

Panu Chaichanavong

panu@stanford.edu

In image compression, pixel-to-pixel correlation is often exploited
to yield higher compression ratio. Differential pulse code modulation
(DPCM) is among the most common predictive coding used to decorrelate
the image pixels. However, this scheme is causal, i.e. it uses only
previously encoded pixels to predict a pixel?s value. Burt and Adelson
[1] introduced a noncausal prediction scheme which obscures this drawback. 

In this project, we implement the classical Burt/Adelson pyramid coder.
According to their algorithm, the only variable is the parameter 'a' in
their filter. We investigate the effect of this parameter and find the
optimum filter kernel.

Important components in Burt/Adelson's so called Laplacian pyramid
structure are filter and interpolator. We propose a nonlinear predictor
by changing these components to improve its performance. The result is
compared to the original scheme.

[1] P. J. Burt and E. H. Adelson, 'The Laplacian pyramid as a compact
code,' IEEE Transactions on Communications, vol. COM-31, pp. 532-540,
Apr. 1983.