Image denoising using tree-based wavelet subband
   correlations and shrinkage

Abstract 

In this paper we describe new methods of denoising images 
which combine wavelet shrinkage with properties related 
to the statistics of quad-trees of wavelet transform 
values for natural images.  They are called Tree-Adapted
Wavelet Shrinkage (TAWS) methods.  The shift-averaged 
version of TAWS produces denoisings which are comparable 
to state of the art denoising methods, such as cycle-spin
thresholding and the cycle-spin version of the Hidden 
Markov Tree method.  The non-shift averaged version of 
TAWS is superior to the classic wavelet shrinkage method,
and fits naturally into a signal compression algorithm. 
These TAWS methods bear some relation to the recently 
proposed Hidden Markov Tree methods, but are deterministic
rather than probabilistic.  They may prove useful in settings
where speed is critical and/or signal compression is needed.