Return to the homepage
Search the complete Wavelet Digest database
Help about the Wavelet Digest mailing list
About the Wavelet Digest
The Wavelet Digest
> Volume 7, Issue 5
Question: Rediscovering Haar wavelets & wavelet packets

Previous :: Next

Author 
Message 
"Dr. A. Avudainayagam" (avudai@acer.iitm.ernet.in) Guest

Posted: Thu Jan 01, 1970 12:59 am Subject: Question: Rediscovering Haar wavelets & wavelet packets




#30 Question: Rediscovering Haar wavelets & wavelet packets
Recently Hugh Casement (Volume 7, Issue 3) attempted a numerical
experiment which resulted in the rediscovery of Haar wavelets! Here we
want to describe our similar efforts in meddling with Haar wavelets
which appears to give better(!) results when the aim is to make a
vector (representing a signal sampled at equal intervals of time)
sparse.
We simply swap the bottom n/2 rows and the top n/2 rows in an
nxn Haar matrix and call as H'.
_ _
 s s 0 0 . . . . . . 
 0 0 s s . . . . . . 
 . . . . . . . . . . 
 . . . . . . . . . . 
 0 0 . . . . . . s s 
H'=  s s 0 0 . . . . . .  , s=1/sqrt(2)
 0 0 s s . . . . . . 
 . . . . . . . . . . 
 . . . . . . . . . . 
 0 0 0 0 . . . . s s 
 
We apply H' in the same way as the Haar matrix on an nx1 vector which
means that we are transforming the details instead of the averages.
For any chosen threshold, H' makes the sampled signal more sparse than
the traditional Haar matrix H. For example: Consider the signal f(t)
= t squared, 0 <= t < 1 sampled at i/128, 0 <= i <= 127. Note that
for threshold value = 0.000001, the H'transformed signal has 16
elements which may be ignored while the conventional Htransformed
signal has no element to be ignored. Similar results are obtained for
other signals given below. Threshold value = 0.000001
_____________________________________________________________________________
  number of elements which can be number of elements which can be 
  ignored in H transformed signal ignored in H' transformed signal 
_____________________________________________________________________________
   
t cubed  1  9 
 sin(t)  0  8 
 cos(t)  0  12 
 exp(t)  0  8 

A similar improvement(!) can also be acheived for compressing matrices (images).
Question: Does this mean H' compresses the signal better than H? 





All times are GMT + 1 Hour

Page 1 of 1 
