Dividing by the attenuation amplifies the noise and the quantization
steps of the dark areas by the same factor. The wavelet mode performs
the division in the pixel domain, then transforms the result and its
per pixel noise variance, propagated through the squared filter taps as
if the samples were uncorrelated, with the undecimated 9/7 wavelet.
Each detail coefficient is Wiener shrunk toward 0 with a signal
variance estimated from a 7x7 box of coefficients at the dilation of
the level, weighted by their precision and counted only when it exceeds
the noise by 3 standard errors of independent coefficients. The lowpass
of the top level averages over large areas and is kept, so the shade of
heavily attenuated areas survives.
Assisted-by: Claude
Michael Niedermayer
5e56cabf83