What Is Signal denoising?

Signal denoising is the process of separating a meaningful signal from unwanted noise that has been added during measurement, transmission, or storage. In any physical measurement, the observed data is a mixture of the true signal and noise contributions from thermal fluctuations, quantization error, interference from neighboring systems, or sensor imperfections. Denoising recovers the underlying signal by exploiting the difference in structure between signal and noise: signals typically exhibit spatial or temporal correlation, periodicity, or compressibility in some transform domain, while additive noise tends to be broadband and uncorrelated across time or frequency. The field draws on statistical estimation theory, digital signal processing, and machine learning, and its methods are applied wherever high-fidelity signal recovery matters.

Classical denoising theory connects to the Wiener filter formulation, which defines the optimal linear estimator for a signal corrupted by stationary Gaussian noise in terms of the ratio of signal and noise power spectra. Non-stationary signals and structured noise violate the Wiener filter's stationarity assumption, motivating the development of adaptive and transform-domain methods.

Noise Models and Signal Representation

Effective denoising begins with characterizing the noise. Additive white Gaussian noise (AWGN) is the canonical model, assumed in most theoretical analyses, where noise at each sample is drawn independently from a zero-mean Gaussian distribution with known variance. Real measurements often involve correlated noise (structured interference from power lines at 50 or 60 Hz), multiplicative noise (speckle in ultrasound), or impulsive noise (transient spikes in industrial environments). The choice of denoising method depends critically on which noise model applies. Similarly, the signal representation matters: a biomedical signal like an ECG is not white in the time domain, and a speech signal is not stationary over long windows; transforms that concentrate the signal energy into a small number of coefficients while spreading noise broadly are the basis of effective denoising strategies.

Filtering-based Denoising

Linear filters such as low-pass, bandpass, and Wiener filters are the simplest denoising tools. A low-pass filter preserves frequency components below a cutoff and attenuates higher frequencies where broadband noise dominates, at the cost of blurring transient features. Adaptive filters adjust their coefficients in real time based on an error signal, allowing them to track changes in noise statistics. The least mean squares (LMS) and recursive least squares (RLS) algorithms are the standard adaptive approaches used in acoustic noise cancellation and biomedical signal processing. Median filters provide robust noise removal for impulsive contamination because they replace each sample with the median of its neighbors, inherently suppressing isolated outliers without convolving the signal with a kernel that spreads spike energy.

Wavelet and Transform-domain Methods

Wavelet denoising, introduced by Donoho and Johnstone in the 1990s, exploits the sparsity of smooth or piecewise-regular signals in the wavelet domain. The procedure is to decompose the noisy signal into wavelet coefficients, apply a thresholding operation (hard or soft) that zeroes or shrinks coefficients below a noise-level threshold, and reconstruct the denoised signal. Because noise distributes its energy roughly uniformly across all wavelet coefficients while the signal concentrates energy in a few large-magnitude ones, thresholding removes the noise while preserving the signal's dominant features. Research on optimal wavelet selection for signal denoising published by PMC at NIH shows that wavelet family choice significantly affects the quality of the recovered signal, and automated selection criteria based on sparsity can outperform hand-tuned defaults. The IEEE survey of wavelet transform methods for signal denoising documents applications in biomedical, audio, and remote-sensing signals. Deep learning approaches based on convolutional neural networks trained on matched signal-noise pairs have extended the wavelet principle to non-stationary, non-Gaussian noise through a data-driven NIH study on wavelet denoising of biomedical signals.

Applications

Signal denoising has applications in a wide range of fields, including:

• Biomedical signal processing for ECG, EEG, and EMG artifact removal
• Audio enhancement and speech intelligibility improvement in hearing aids and communications
• Image and video processing to reduce sensor noise and compression artifacts
• Seismic data processing for oil exploration and earthquake monitoring
• Radar and sonar signal recovery to improve target detection in cluttered environments