next up previous contents index
Next: References Up: Denoising Previous: Asymptotics

The selection algorithm

We chose to view a signal f as being noisy or incoherent relative to the basis if its entropy is of the same order of magnitude as giving a poor compression rate , leading us to the following method.

Start with a signal f of length N, find the best basis in each library and select among them the basis minimizing . Reorder the coefficients in decreasing order where for are all below a precision threshold (say 0.1% of energy). Then decompose

We will say that is coherent and is incoherent if

The threshold is chosen to determine if the compression of using is unacceptably bad. We proceed by testing until we reach M for which

or

We can now consider as a new signal for which we repeat the decomposition, i.e. pick a best basis and decompose

iterating a fixed number of times or stopping whenever no new coherent part is obtained.

The Xwpl program automatically determines the threshold using heuristics on the coefficient curve.



Fazal Majid
Fri Jan 27 11:23:48 MET 1995