castanyes blaves

Wavelets stuff I got buried somewhere else

Meeting with B.A:

It went really well.

About which filter to use, he said that Daubechies is nice because of

the compact support: you know where are the border artifacts in the

signal.

About the different levels of Daubechies:

a0 + a1x + a2x^2 + a3x^3 + ...

D1 - Haar, is ok to remove the constant from the data (a0). So you

remove the constant and have the data normalised in the detail.

And there is no border effect in Haar if signal is 2^n.

D2 - To remove linear trends.

D3 - To remove low frequency trends.

D4 and so on is to de-correlate different point in the signal.

One of the approaches for the project (p1) would be:

1) Characterize the globals of the signal, i.e., have a global view of

how is the signal.

2) Detect the edges in the signal:

-- Simple continuous 'g1' can be used for that: derivative of the signal

at a certain scale. Or also 'g2'.

About the denoising, we have to take into account that if we use a

simple threshold, the coefficient of the threshold at each level has

to be in accordance with Donoho's formulae.

We tested the use of continuous wavelet decomposition with cwtd 1 8 6

'g2'.

They are also working on that but with data of gene density.

We are going to be in touch to collaborate on the issue in a near

future (a couple of months).

It might be worth to talk with P.L., although he is now in the

computer department in UCam.

posted by avilella  # 2:04 PM