Skip to content

Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images

Why this mattered

Before this paper, image restoration was often treated as a signal-processing problem centered on filters, linear inverse methods, or local heuristic smoothing. Geman and Geman reframed it as probabilistic inference over a structured image: an image could be modeled as a Markov random field, equivalently as a Gibbs distribution with an explicit energy function encoding preferences for regions, boundaries, noise, blur, and other degradation processes. This shifted the central object from an ad hoc restoration rule to a posterior distribution over possible images. The restored image became the maximum a posteriori state of that distribution.

The paper also made a difficult inference problem operational. By importing simulated annealing from statistical mechanics, it gave a principled algorithm for seeking low-energy, high-posterior-probability images, with convergence guarantees under gradual temperature reduction. That mattered because it showed that complex visual priors, including discontinuities such as edges, could be combined with observation models in a unified Bayesian framework rather than being bolted on after smoothing. It made possible a generation of energy-based methods for denoising, deblurring, segmentation, texture modeling, stereo, and other vision tasks where local interactions produce global structure.

Its influence extended well beyond image restoration. The paper helped establish the language of graphical models, energy functions, posterior inference, and MAP estimation as standard tools in computer vision. Later breakthroughs such as graph cuts, belief propagation, variational inference, conditional random fields, and modern energy-based models all inherit part of this conceptual program: define a structured probability or energy model, then solve the corresponding inference problem. The specific annealing procedure was often replaced by faster approximations, but the paradigm it made convincing endured.

Abstract

We make an analogy between images and statistical mechanics systems. Pixel gray levels and the presence and orientation of edges are viewed as states of atoms or molecules in a lattice-like physical system. The assignment of an energy function in the physical system determines its Gibbs distribution. Because of the Gibbs distribution, Markov random field (MRF) equivalence, this assignment also determines an MRF image model. The energy function is a more convenient and natural mechanism for embodying picture attributes than are the local characteristics of the MRF. For a range of degradation mechanisms, including blurring, nonlinear deformations, and multiplicative or additive noise, the posterior distribution is an MRF with a structure akin to the image model. By the analogy, the posterior distribution defines another (imaginary) physical system. Gradual temperature reduction in the physical system isolates low energy states (annealing''), or what is the same thing, the most probable states under the Gibbs distribution. The analogous operation under the posterior distribution yields the maximum a posteriori (MAP) estimate of the image given the degraded observations. The result is a highly parallelrelaxation'' algorithm for MAP estimation. We establish convergence properties of the algorithm and we experiment with some simple pictures, for which good restorations are obtained at low signal-to-noise ratios.

Sources