
Scalespace peak picking

This function looks for peaks in the data using scalespace theory.
input :
* V : data, a vector
* select : either:
 select >1 : the number of peaks to detect
 0<select<1 : the threshold to apply for finding peaks
the closer to 1, the less peaks, the closer to 0, the more peaks
* display : whether or not to display a figure for the results. 0 by
default
* ... and that's all ! that's the cool thing about the algorithm =)
outputs :
* peaks : indices of the peaks
* criterion : the value of the computed criterion. Same
length as V and giving for each point a high value if
this point is likely to be a peak
The algorithm goes as follows:
1°) set a smoothing horizon, initially 1;
2°) smooth the data using this horizon
3°) find local extrema of this smoothed data
4°) for each of these local extrema, link it to a local extremum found in
the last iteration. (initially just keep them all) and increment the
corresponding criterion using current scale. The
rationale is that a trajectory surviving such smoothing is an important
peak
5°) Iterate to step 2°) using a larger horizon.
At the end, we keep the points with the largest criterion as peaks.
I don't know if that kind of algorithm has already been published
somewhere, I coded it myself and it works pretty nice, so.. enjoy !
If you find it useful, please mention it in your studies =)
running time should be decent, although intrinsically higher than
findpeaks. For vectors of length up to, say, 10 000, it should be nice.
Above, it may be worth it though.

(c) Antoine Liutkus, 2013

