I have the following function obtained from averaging 2d camera image over one axis, where detection of aligned objects is desired:

function to be analyzed

I need to detect (roughly speaking) locations and widths of the troughs -- sometimes and object will be missing, thus the trough will be much wider. The approach should have some resilience WRT possible lighting variations and other noise; this, and also a desire to use something new and useful) prevents me from going the naive way of thresholding and clustering or something similar.

I was looking at using wavelets: using ideal trough (Haar?) as the mother wavelet function, and the analysis would give its translation and scale at troughs in the signal.

I have no prior experience with wavelets. My background is numerics (variational analysis/FEM, particle systems programming), c++ and python, but I am quite new to signal processing (and its terminology).

What would be the most suitable thing to do? Suggestions, pointers to articles, books, or, best, code examples are much appreciated.


I finally used continuous wavelet transform (as implemented in scipy) where the result local maxima (in yellow) show both location (x-axis) and width (y-axis) of the trough.

import numpy as np
from scipy import signal
# scale the ricker (mexican hat) function to 10-60px width

enter image description here


2 Answers 2


I have a solution for your problem other than wavelet transforms that could be worth trying. Here are the steps :

  1. Apply AGC & normalise the time series data. This will ensure that even in different lighting conditions etc and you will have relative similar peaks and troughs amplitudes.
  2. Choose a wavelet that best represents your individual event. Eg : looks like inverted Haar could model your troughs well. Or cut a representative part of the normalised time series (your data). Say this is your reference wavelet W.
  3. Correlate the reference Wavelet W with the data (time series). The values of correlation coefficients will give you a measure of spread of the troughs / peaks. The largest value will give the location.

Here is small code python code: if x is your time series data W is reference wavelet :

import numpy as np 
corr_coeff= np.correlate(x,W,'same')
  1. Though this will still require some thresholding and tests , normalisation should make it simpler.

Reference : https://en.wikipedia.org/wiki/Cross-correlation


A wavelet analysis or wavelet transform will give you scale(frequency) vs time vs amplitude. I am not sure how that will help you with your problem. Here is an algorithm for wavelet transform :

  1. Choose your wavelet : Haar , Molet etc ..say $ \psi (n) $
  2. Choose Scale : the scale should be a multiple of smallest time interval dt.
  3. Calculate the FFT of wavelet - $\Psi_0 (s\omega _k)$
  4. Normalise FFT of the Wavelet - $ \Psi (s\omega_k)=(2\pi s / \delta t)\Psi_0 (s\omega _k)$
  5. Find FFT of time series x(t) = $ X_k = 1/N \sum_{k=0}^{N-1} x(n) e^{-2\pi i k n/N } $
  6. Then Wavelet transform is : $ W(s)= \sum_{k=0}^{N-1} X_k\Psi^* (s\omega_k) e^ {i\omega_k n \delta t} $ ...where * is complex conjugate Generally the Fourier transform of the wavelet should be known analytically.

Few good sources :

  1. http://paos.colorado.edu/research/wavelets/bams_79_01_0061.pdf

  2. http://users.rowan.edu/~polikar/WAVELETS/WTpart1.html


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.