The Wavelet Digest Homepage
Return to the homepage
Help about the Wavelet Digest mailing list
About the Wavelet Digest
This website was frozen December 31, 2012
The Digest The Community
 Latest Issue  Back Issues  Events  Gallery
The Wavelet Digest
   -> Volume 6, Issue 9

Answer: Shiftability (WD 6.8 #25)
images/spacer.gifimages/spacer.gif Reply into Digest
Previous :: Next  
Author Message (Andrew Dorrell)

PostPosted: Wed Aug 13, 1997 7:41 am    
Subject: Answer: Shiftability (WD 6.8 #25)
Reply with quote

#13 Answer: Shiftability (WD 6.8 #25)

<From the Question>

There they have over come an old problem of the shift invariance of
the wavelet transform by not down sampling while taking the wavelet
transform.They have showed with examples that it works.

I would like to know why this works and will it always work? If you
have any suggestions or any reference materials,can you please
forward them to my address.

Yes it will always work. The reason is simply that shift variance is
introduced because Nyquist sampling is violated in each of the
(wavelet decomposed) subbands - wavelet are not ideal filters after
all! By not downsampling the problem is avioded. There are in fact
many informative articles on this subject. Daubechies talks about
creating tight frames in chapter 3 of her definative book. Although
not explicit, the motivation is shift invariance. Chui has also
published a number of papers on maintaining tight frames at certain
oversampling rates. A well explained work. The 'a-trous' algorithm
is in fact a fast implementation of a wavelet transform with no
downsampling - used because of its shift invariance. Eero Simoncelli
(look up his web page) also has a number of papers specifically
addressing the issue of shiftability and developing jointly shiftabl,
and overcomplete, wavelet tight frames. More recently a number of
authors have examined other methods of achieving shiftability - some
are as simple as deciding, on a level by level basis, whether to
retain odd or even samples. Anyway the motivation for using these
other thechniques in preference to the a-trous (not downsampled)
wavelet transform is thatthey retain some of the compactness
properties, and reduced computational complexity of orthogonal

Whether some are applicable may still depend on the data you are working

Mr Andrew Dorrell
School of Electrical Engineering
University of Technology, Sydney

[Updated answer]
All times are GMT + 1 Hour
Page 1 of 1


This page was created in 0.026065 seconds : 18 queries executed : GZIP compression disabled