1
$\begingroup$

My mathematical class task is to de-noise a function using wavelet transform. I am to select a function $f(x)$ and noise function with zero-mean $n(x)$. I am to add noise like this: $$f_{noise}(x) = f(x) + \epsilon n(x)$$ The thing is, I should perform de-noising process not just by using thresholding or a similar technique, but taking into account the available information about the noise function. If I were to do de-noising using Foruier transform, I would null frequencies corresponding to noise function, but wavelets are different since detailing coefficients aren't actually frequencies, so I fail to come up with a solution. So, my question is: how to denoise a function with wavelet transform relying on the fact that $n(x)$ is known?

$\endgroup$
  • $\begingroup$ Can you explain what information you know about the noise function? Surely you don't simply know $n(x)$ exactly, because then you could just subtract it from $f_\text{noise}(x)$ and be done... $\endgroup$ – Rahul Dec 24 '13 at 15:20
  • $\begingroup$ That's true! Let's say I know, that this is a Gaussian noise $\endgroup$ – Dmitry Dec 24 '13 at 15:25
  • $\begingroup$ I don't have an answer, but I'd like to point out another SE site, Signal Processing. $\endgroup$ – Post No Bulls Dec 25 '13 at 7:03
1
$\begingroup$

I assume the signal is being sampled on a discrete set $\{ x_i \}$, say evenly spaced on $[0,1]$. Fix a model space $X$ for the signal $f$, e.g. Sobolev, Besov, etc.

Typically, one would perform a discrete wavelet transform (where the filter is given by some chosen wavelet basis for $X$) and threshold the wavelet coefficients using some criterion involving $\epsilon$. For example, hard thresholding means setting any wavelet coefficients $ < \epsilon$ to $0$. Then you invert the thresholded coefficients. This gives the estimate for $f$.

The mathematical and statistical justification of this procedure lies in the fact that

  1. Wavelets can describe signals parsimoniously.
  2. Wavelets form an unconditional basis for $X$; so thresholding preserves smoothness.
$\endgroup$
  • $\begingroup$ What do you mean by 'invert thresholding coefficients'? If, say, detailing coefficients are [1 2 1 3 4] and the threshold is 2 then the resulting detailing coefficients would be [0 2 0 3 4] and isn't it enough? $\endgroup$ – Dmitry Dec 25 '13 at 17:21
  • $\begingroup$ By "invert", I mean applying the inverse DWT to the thresholded coefficients $[0\; 2\; 0\; 3\; 4]$. $\endgroup$ – Michael Dec 26 '13 at 22:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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