Besov regularization, thresholding and wavelets for smoothing data (Q4385289)

If \(x\) represents noisy data and \(Wx\) its wavelet transform, then smoothed data can be obtained via \(y=s(x)=W^{-1}r(Wx)\), where \(r\) is a (usually nonlinear) thresholding or some other smoothing operation on the wavelet coefficients. In this paper, it is shown that most known techniques of this type can be formulated as a regularization problem where one wants to minimize \(\| x-y\| +\lambda g(x)\) where \(\lambda>0\) is a smoothing parameter and \(g(x)\) some smoothing function. The \(g(x)\) can be chosen so as to recover hard and soft thresholding or other smoothing techniques. When orthogonal wavelets are used, this optimization problem is easily expressed in terms of the wavelet coefficients and in such cases explicit solutions are obtainable. This paper introduces Besov regularization as an alternative, where \(g(x)\) is some Besov norm. Here the solution is not explicit, but relatively simple to find. Also level dependent smoothing is possible, choosing different parameters per resolution level. Numerical experiments illustrate the performance of the different methods.