Noise reduction is the family of techniques that attenuate unwanted random or structured disturbances in a signal while preserving the underlying information of interest. It ranges from classical linear and spectral filtering to statistical estimators and learned denoising models that infer clean signals from noisy observations. Effective noise reduction improves downstream perception, measurement and machine-learning tasks by raising the signal-to-noise ratio.
Overview
- Noise reduction trades off suppression of disturbances against preservation of genuine detail.
- Classical methods include moving-average, Wiener and median filters and spectral subtraction.
- Learned denoisers, including autoencoders and diffusion priors, exploit data statistics for stronger results.
Key aspects
- Modelling noise as additive, multiplicative or structured interference.
- Linear filtering versus non-linear and edge-preserving filters.
- Statistical estimation under known or estimated noise distributions.
- Data-driven denoising trained on clean-noisy pairs or self-supervised.
Applications
- Audio enhancement and speech denoising.
- Medical and astronomical image cleanup.
- Sensor pre-processing for robotics and instrumentation.