and thanks for your reply.
Well, using the symmetric haar h = [1/16 1/4 3/8 1/4 1/16] I do not have lag effects and coefficents are in phase perfectly but I get border effects that's unuseful to be trained with neural nets.
The same unuseful training is because of lag effects using the 1/2 1/2.
I wonder how do they do to get a perfect overlap of the coefficents as in figure 2 with a redundant haar a trous!
The trick is just to get a wavelets that is both lag-free and without border effect, but I have been unable to code.
2005/6/3, Jeff Winter <j...@aeroflex.com>:
If you ignore the lag, does the result look ok?
Non-symmetric filters introduce phase shifts. The effect of a phase shift
is to displace events along the time axis; that is, if a maximum in the
original series occurs at time t0, the phase shifted maximum occurs at time
t0 + delta, where delta may be positive or negative.
So while the non-symmetric filter is important to detect edge boundaries in
this application, I think the delay is unavoidable. However, from the look
of this paper a lot of the results show the effects of de-noising the
signal? The (1/2, 1/2) filter coefficients are a simple averaging filter
that fit in with the de-noising hypothesis.
Snr DSP Engineer