Wavelet Shrinkage: From Sparsity and Robust Testing to Smooth Adaptation
Wavelet transforms are said to be sparse in that they represent smooth and piecewise regular signals by coefficients that are mostly small except for a few that are significantly large. The WaveShrink or wavelet shrinkage estimators introduced by Donoho and Johnstone in their seminal work exploit this sparsity to estimate or reconstruct a deterministic function from the observation of its samples corrupted by independent and additive white Gaussian noise AWGN. After a brief survey on wavelet shrinkage, this chapter presents several theoretical results, from which we derive new adaptable WaveShrink estimators that overcome the limitations of standard ones, have an explicit close form and can apply to any wavelet transform (orthogonal, redundant, multi-wavelets, complex wavelets, among others) and to a large class of estimation problems. These WaveShrink estimators do not induce additional computational cost that could depend on the application and let the user choose freely the wavelet transform suitable for a given application, in contrast to all parametric methods and also in contrast to some non-parametric methods. In addition, our estimators can be adapted to each decomposition level thanks to known properties of the wavelet transform. The two theoretical results on which our WaveShrink estimators rely are, fist, a new measure of sparsity for sequences of noisy random signals and, second, the construction of a family of smooth shrinkage functions, the so-called Smooth Sigmoid Based Shrinkage (SSBS) functions. The measure of sparsity is based on recent results in non-parametric statistics for the detection of signals with unknown distributions and unknown probabilities of presence in independent AWGN. The SSBS functions allow for a flexible control of the shrinkage thanks to parameters that directly relate to the attenuation wanted for the small, median and large coefficients. The relevance of the approach is illustrated in image denoising, a typical application field for WaveShrink estimators.
KeywordsAdditive White Gaussian Noise Decomposition Level Image Denoising Constant False Alarm Rate Wavelet Shrinkage
Unable to display preview. Download preview PDF.
- 3.Atto, A.M., Pastor, D., Mercier, G.: Smooth Sigmoid Wavelet Shrinkage for Non-Parametric Estimation. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Las Vegas, NV, USA (2008)Google Scholar
- 6.Coifman, R.R., Donoho, D.L.: Translation invariant de-noising. Lecture Notes in Statistics: Wavelets and Statistics, 125–150, Springer, New York (1995)Google Scholar
- 12.Holschneider, M., Kronland-Martinet, R., Morlet, J., Tchamitchian, P.: A Real-Time Algorithm for Signal Analysis with the Help of the Wavelet Transform. In: Wavelets: Time-Frequency Methods and Phase-space. Springer, Berlin (1989)Google Scholar
- 22.Minkler, G., Minkler, J.: The Principles of Automatic Radar Detection in Clutter, CFAR. Magellan Book Company, Baltimore, MD (1990)Google Scholar
- 25.Pastor, D., Socheleau, F.-X., Aïssa-El-Bey, A.: Sparsity Hypotheses for Robust Estimation of the Noise Standard Deviation in Various Signal Processing Applications. Second International Workshop on Signal Processing with Adaptive Sparse Structured Representations, SPARS’09, Saint-Malo, France (2009)Google Scholar
- 32.Socheleau, F.-X., Pastor, D., Aïssa-El-Bey, A., Houcke, S.: Blind Noise Variance Estimation for OFDMA Signals. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Taipei, Taiwan (2009)Google Scholar