What Is Deconvolution?

Deconvolution is the process of recovering an original signal or image from an observed signal that has been distorted by convolution with a known or unknown blurring function. In physical measurement systems, the observed output is the convolution of the true signal with a system's impulse response, also called the point spread function (PSF) in imaging contexts. Deconvolution inverts this operation to estimate the underlying signal. It is an ill-posed inverse problem: small amounts of measurement noise can produce large errors in the recovered signal, so regularization or prior information is essential for stable solutions.

The field draws from linear systems theory, functional analysis, and computational methods. The mathematical framework for convolution and its inversion is rooted in Fourier analysis, where convolution in the time or spatial domain corresponds to multiplication in the frequency domain. Wiener, working in the 1940s on optimal linear filtering, derived the Wiener filter as the minimum mean-squared-error deconvolution estimator under additive Gaussian noise, and this result remains a standard reference solution against which more sophisticated methods are benchmarked.

Blind Deconvolution

When the PSF is unknown, the problem becomes blind deconvolution: both the original signal and the blurring function must be estimated simultaneously from the observed data alone. This is inherently underdetermined, so algorithms impose constraints such as sparsity in the signal, non-negativity, band-limitedness, or parametric models for the PSF shape. The IEEE Xplore paper on blind image deconvolution surveys early algorithms that alternate between estimating the PSF and estimating the latent image in an expectation-maximization style iteration. Modern deep learning approaches train convolutional neural networks on large paired datasets of blurred and sharp images to learn a mapping that implicitly encodes both image priors and PSF models, achieving restoration quality that surpasses classical methods on natural images while requiring no explicit PSF estimation at inference time.

Blind deconvolution in one-dimensional signal processing appears in seismology, where the earth's reflectivity sequence is the unknown signal and the seismic wavelet plays the role of the PSF. Predictive deconvolution in seismic processing, based on the Wiener-Levinson algorithm, exploits the autocorrelation structure of the seismic trace to estimate and remove the wavelet, revealing the reflectivity series that encodes subsurface structure.

Signal Restoration and Regularization

Non-blind deconvolution, where the PSF is measured or specified, remains practically important in microscopy, astronomy, and spectroscopy. Naive inversion in the frequency domain amplifies high-frequency noise catastrophically, because the PSF typically attenuates high spatial frequencies and division by small numbers magnifies noise. The Wiener filter addresses this by incorporating a noise-to-signal power ratio that suppresses restoration at frequencies where noise dominates. Tikhonov regularization adds a penalty on the energy or smoothness of the reconstructed signal. The NIST guidelines on measurement uncertainty and signal processing provide context for quantifying how measurement noise propagates through deconvolution and how regularization parameter selection balances bias against variance in the estimated output.

Iterative algorithms including Richardson-Lucy deconvolution, which maximizes the Poisson log-likelihood and is standard in fluorescence microscopy, and the conjugate gradient method applied to least-squares formulations provide finer control over regularization than closed-form Wiener solutions. The review of deconvolution algorithms from Evident Scientific describes nearest-neighbor, constrained iterative, and blind methods as applied to fluorescence and confocal microscopy, where the PSF is approximately known from optical parameters.

Applications

Deconvolution has applications in a wide range of fields, including:

• Astronomy and radio telescope image reconstruction
• Fluorescence and confocal microscopy for biological imaging
• Seismic signal processing for subsurface geological mapping
• Medical imaging, including CT and MRI image enhancement
• Spectroscopy, for recovering peak shapes from instrument broadening