Wavelet Signal Denoising (Soft Thresholding)
Also known as: Wavelet Shrinkage, Donoho-Johnstone Denoising, Soft Thresholding Denoising, Sinyal Gürültü Giderme
Wavelet signal denoising, introduced by David Donoho in 1995, is a non-parametric technique that removes noise from one-dimensional or multidimensional signals by decomposing them into wavelet coefficients, suppressing small coefficients that likely represent noise via a soft-thresholding operator, and reconstructing a smooth estimate. It is widely used in biomedical signal processing, geophysics, audio engineering, and image analysis where the underlying signal is assumed to be sparse or piecewise smooth.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use wavelet soft-thresholding when your signal is corrupted by additive white Gaussian noise and you expect the true signal to be piecewise smooth, sparse in a wavelet basis, or to contain transient features at multiple scales. It is particularly effective for biomedical waveforms (EEG, ECG), seismic traces, and audio signals. Assumptions include stationarity of the noise and approximate sparsity of the signal. It is less appropriate for colored noise, non-stationary noise variance, or signals with purely sinusoidal structure where Fourier-based methods are preferable.
Strengths & limitations
- Provides near-minimax optimal mean squared error over broad classes of smooth and piecewise-smooth functions.
- Adaptive to local signal features at multiple scales, preserving edges and transients that global Fourier methods smear.
- Universal threshold requires no prior knowledge of the signal; it is computed directly from the data.
- Computationally efficient via the fast wavelet transform, running in O(n) time.
- Assumes additive white Gaussian noise; performance degrades under colored or non-stationary noise.
- The universal threshold can over-smooth signals with low sparsity in the wavelet domain.
- Results depend on the choice of wavelet family and decomposition depth, which require domain knowledge.
- Boundary effects in the wavelet transform can introduce artifacts near signal endpoints.
Frequently asked
What is the difference between soft and hard thresholding?
Hard thresholding sets coefficients below the threshold to zero and leaves the rest unchanged, creating a discontinuous operator that can introduce ringing artifacts. Soft thresholding additionally shrinks surviving coefficients toward zero by the threshold amount, producing a continuous operator with smoother reconstructions at the cost of a small bias on large coefficients.
How is the universal threshold λ = σ√(2 ln n) derived?
The threshold follows from the extreme-value theory of Gaussian random variables. For n independent standard Gaussian variables, the expected maximum grows as √(2 ln n). Setting λ at this level ensures that, with probability tending to one, all pure-noise coefficients fall below the threshold, while the bias introduced to signal coefficients is bounded within a factor of log n of the oracle risk.
When should I prefer EMD or VMD over wavelet denoising?
Empirical Mode Decomposition and Variational Mode Decomposition are preferable when the signal is non-stationary and non-linear, when no a priori wavelet basis fits the signal structure, or when you need data-adaptive oscillatory modes rather than fixed basis functions. Wavelet denoising has stronger theoretical guarantees and is faster, making it the default choice when signal sparsity in a standard wavelet basis is plausible.
Sources
- Donoho, D. L. (1995). De-noising by soft-thresholding. IEEE Transactions on Information Theory, 41(3), 613–627. DOI: 10.1109/18.382009 ↗
How to cite this page
ScholarGate. (2026, June 2). Wavelet Signal Denoising (Soft Thresholding). ScholarGate. https://scholargate.app/en/signal-processing/signal-denoising
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Empirical Mode DecompositionSignal Processing↔ compare
- Fourier TransformSignal Processing↔ compare
- Variational Mode DecompositionSignal Processing↔ compare