Denoisers

Denoisers are algorithms designed to remove or reduce noise from a signal, including images, video, audio, or other data, with the aim of recovering the underlying clean signal. They are applied across imaging, communications, and audio processing and are typically framed as inverse problems, where the observed data equals the true signal plus noise.

Traditional denoising methods rely on assumptions about the signal and noise. Spatial filters such as mean or median filters are simple but often blur details. More advanced classical approaches include non-local means, BM3D, wavelet thresholding, and Wiener filtering. These methods exploit redundancy and sparsity in suitable representations to separate signal structure from noise.

Transform-domain and model-based methods use priors such as sparsity, smoothness, or low-rank structure. Total variation regularization and sparse coding are examples, often solved by optimization techniques. In many cases, the noise model is assumed Gaussian, Poisson, or mixed, and methods may be tuned to a known noise level.

Data-driven denoisers use machine learning to learn mappings from noisy inputs to clean outputs. Supervised neural networks, such as convolutional denoisers, can outperform traditional methods when training data match the target domain. Unsupervised or self-supervised variants, including blind-spot and Noise2X approaches, aim to avoid requiring clean targets. More recently, diffusion-based and other deep generative models have been explored for denoising.

Common applications include digital photography, medical and astronomical imaging, video restoration, and audio enhancement. Evaluation typically uses metrics such as PSNR or SSIM for images, along with perceptual assessments. Challenges include preserving fine details, avoiding artifacts, and generalizing across noise types and levels.