Restoration is inference about an image that was never observed
An image-restoration problem begins with an asymmetry. The scene of interest, denoted by an intensity field F, is hidden. A recording device blurs it, may transform the blurred values nonlinearly, and introduces sensor noise. Only the degraded field G is available. Many different originals can lead to observations that are similar or even indistinguishable. Inverting the recording equation therefore requires more than undoing a matrix multiplication: it requires a defensible account of which originals were plausible before the data arrived.
Geman and Geman write a deliberately broad degradation scheme. A shift-invariant point-spread operator H forms a blurred image; a memoryless transformation φ may act on those values; an independent noise field N is then combined through an operation that could be addition or multiplication. The point is not that every camera literally follows one notation. The point is modularity. Optics, response nonlinearity and sensor error belong in the likelihood, while spatial expectations about the unobserved image belong in the prior.
The paper’s central move is to represent both sides as energies. A Gaussian additive-noise likelihood, for example, penalises squared disagreement between the recorded value and the image predicted by the forward model. A spatial prior penalises patterns judged unlikely, such as isolated grey-level changes inside a region. Taking negative logarithms converts multiplication of probability factors into addition of energy terms. Restoration becomes the search for a low point on a landscape whose shape records both physical evidence and structural expectation.
This is already a departure from a merely linear filter. The estimate at one pixel can depend on a nonlocal configuration through a chain of local interactions; edges can be represented by latent variables; and the final rule is nonlinear. The 1984 article explicitly describes its models and estimates as noncausal and nonlinear, inspired by lattice systems in statistical physics rather than by extending one-dimensional smoothing formulas into a second dimension.
Thought experiment 1 · One recording, two priors
Setup. A noisy 8 × 8 observation contains a faint diagonal streak. Single change. Analyst A uses a prior favouring compact regions; analyst B uses one favouring thin connected curves. The likelihood is identical. Result. A may erase the streak as noise while B restores it as structure. The disagreement is not a failure of Bayes’s rule; it exposes the different image classes encoded before inference. A restoration must therefore report its prior, not only its optimiser.
Worked example 1 · Data fit alone is underdetermined
Suppose a single observed value is g = 0.55 and the noise standard deviation is 0.20. Candidate intensity f = 0.4 has squared-error energy (0.55 − 0.4)²/(2 × 0.2²) = 0.28125. Candidate f = 0.7 has the same energy because it is equally far from g. The likelihood cannot choose. If the left and right neighbours are near 0.4, a continuity prior adds little energy to f = 0.4 and more to f = 0.7. The posterior choice comes from evidence plus context, not evidence alone.
A Markov field turns distant organisation into local conditionals
Let the image sites form a finite lattice. Each site carries a grey level from a finite set. A neighbourhood system says which sites directly interact: perhaps the four orthogonal neighbours, perhaps all eight surrounding sites, or a richer graph that also includes latent variables. The Markov property then makes a precise conditional-independence claim. Once the neighbours of site s are known, the value at s is conditionally independent of every other site.
This is local conditional dependence, not marginal independence. Two distant pixels can remain strongly associated because influence travels along paths through the graph. That distinction explains why a field built from small neighbourhoods can display broad regions, long boundaries and texture. It also explains the computational opportunity: to update one site, the sampler need not reconsider the entire image.
The Gibbs representation supplies the joint law. Assign a potential to every clique, where a clique is a set of mutually neighbouring sites. Sum those potentials to obtain the energy U(f). Configurations with lower energy receive exponentially greater probability, normalised by a partition function Z. For finite positive fields, the MRF and Gibbs descriptions are equivalent. Geman and Geman rely on both faces of that equivalence: local conditionals make computation feasible, while clique potentials make structural modelling intelligible.
Temperature T scales discrimination among energies. At a high temperature, even configurations well above the minimum retain appreciable probability. At a low temperature, an energy difference is magnified by division through a small T. The normalising constant Z may require summing over astronomically many complete images and is therefore generally intractable. Crucially, it cancels in a one-site conditional ratio. The field can be sampled locally even when its joint normaliser cannot be computed directly.
Worked example 2 · A local conditional without Z
A binary site has four neighbours: three are 0 and one is 1. Give each disagreeing bond energy β = 0.8 and ignore data for the moment. Setting the site to 0 creates one disagreement, so E₀ = 0.8. Setting it to 1 creates three, so E₁ = 2.4. At T = 1, Pr(Fₛ=1 | neighbours) = exp(−2.4)/[exp(−0.8)+exp(−2.4)] ≈ 0.168. The unknown partition function for every complete image never appears. The update needs only two local energies.
Thought experiment 2 · Local graph, long-range consequence
Setup. Flip the upper-left pixel of a nearly uniform field. Single change. Run one update only, then many complete sweeps. Result. The first decision consults only nearby pixels, yet later neighbours react, their neighbours react, and the influence can propagate across the lattice. “Markov” does not mean “short-sighted output”; it means that the mechanism factors through local conditional steps.
Smoothing needs permission to stop at an edge
A simple pairwise prior rewards neighbouring pixels for sharing a grey level. That is useful inside a region and destructive at a boundary. Increase the smoothness penalty and noise disappears, but so do narrow objects and sharp transitions. Decrease it and boundaries survive, but isolated corruptions remain. The paper’s hierarchical answer is to introduce a second field, L, on a dual lattice of sites between pixels.
Each line variable indicates whether the ordinary intensity bond is intact or broken. When no line element is present, different neighbouring intensities incur their usual penalty. When the bond is broken, that intensity disagreement can cost zero. The line process has its own clique potentials: isolated fragments, parallel pieces and sharp turns can be expensive; continuous or gently turning boundaries can be cheaper. Thus the model does not merely permit every discontinuity. It prefers discontinuities that assemble into plausible contours.
The observable intensity field is the marginal of the joint field X = (F,L). Marginalising over possible lines can induce complicated and effectively long-range dependence among intensities, even though the joint F–L graph remains local. This is a deep reason for the hierarchy. A richer visible model is obtained without surrendering local computation. Later graphical-model practice would make this pattern familiar: introduce latent structure so the expanded model factors simply.
Worked example 3 · When a broken bond wins
Consider adjacent observed pixels near 0 and 2. Candidate intensities (0,2) fit the data exactly but pay a smoothness cost of 2 when the bond is intact. Candidate (1,1) pays no smoothness cost but two units of data energy, one for each altered pixel. They tie. If a line element costs 0.7 and turns off the bond, the joint state (0,2,line-on) costs 0.7 and beats both alternatives. If isolated line elements cost 3 instead, the same discontinuity is suppressed. Edge preservation therefore depends on a model of lines, not a universal command to keep every jump.
Thought experiment 3 · Remove the line layer
Setup. Restore the same hand-drawn shape from the same noisy recording with the same intensity levels. Single change. Remove L while retaining a pairwise blob-like smoothness prior. Result. The paper’s Group 2 comparison shows the characteristic tradeoff: a pure blob process cleans interiors but damages geometry; adjoining a line process permits clean regions and sharp boundaries together. The latent layer changes the hypothesis class, not merely the optimisation speed.
The posterior inherits locality from likelihood and prior
Bayes’s rule says the posterior is proportional to likelihood times prior. In energy form, the posterior energy is the sum of a degradation term and an image term. If the forward mechanism is spatially local;or can be represented with a manageable extended neighbourhood;the posterior remains a Gibbs field with a related graph. This preservation is the paper’s pivotal computational observation. Nonlinear response and multiplicative noise need not destroy the method; they alter the local data potentials.
For additive independent Gaussian noise without blur, the data term decomposes by pixel: each candidate intensity pays a squared residual. A pairwise prior then adds neighbour costs; a line prior adds costs for contour configurations. Blur enlarges the effective neighbourhood because one observation depends on several original pixels, but the method remains usable when that footprint stays reasonable. “Local” is consequently an engineering property of the chosen factorisation, not a claim that cameras affect only one pixel at a time.
The lowest posterior energy corresponds to a maximum a posteriori estimate. This is a joint mode of F and L when the hierarchy is retained. It is not the same as the pixelwise posterior mean, a draw from the posterior, or a guarantee of minimum expected error under every loss. The article notes that temporal ergodicity could also support minimum mean-square estimation through posterior averages, but its restoration experiments pursue MAP configurations.
Worked example 4 · Two candidates, three accounts
Candidate A fits the observed pixels with data energy 8, has intensity energy 1 and line energy 5, total 14. Candidate B softens the data slightly;energy 10;but forms smoother regions with intensity energy 2 and a coherent contour costing 1, total 13. MAP chooses B. Calling B “closer to the truth” would be too strong: it is more probable under this likelihood–prior combination. If line fragments were cheaper or the noise variance changed, the ranking could reverse.
The Gibbs sampler refreshes one site from the field itself
Directly enumerating an image is hopeless. With L possible intensities on an m × m lattice, there are L raised to m² complete intensity configurations before line states are counted. The Gibbs sampler avoids that combinatorial list. Choose a site, hold every other site at its current value, calculate the site’s full conditional distribution, draw a replacement, and repeat. Because the field is Markov, only the factors touching that site need be recomputed.
A draw can keep the existing value or replace it with a higher-energy value. That randomness is essential. A deterministic rule that always chooses the locally best value can become trapped in a configuration that is stable under every one-site change but is not globally best. At a positive temperature, occasional uphill moves allow the chain to cross local barriers. The process is called relaxation because repeated local replacements move the system towards its equilibrium distribution.
Theorem A concerns a fixed target temperature. If every site is visited infinitely often and the field is positive on its finite configuration space, the distribution of the current configuration converges to the Gibbs distribution π, regardless of the starting image. The claim is about the probability law of X(t), not independence of successive images. Consecutive states usually differ at no more than one site and can be highly correlated.
Theorem C concerns time averages at a fixed temperature. Under a bounded-gap visiting condition;every site occurs within each sufficiently long block;the average of any function Y(X(t)) converges almost surely to its expectation under π. This connects local simulation to quantities that are hard to sum exactly. In modern language, it is an ergodic theorem for the systematic-scan process considered by the authors.
Worked example 5 · Temperature changes willingness to disagree
Let two candidate values have local energies E₀ = 0.8 and E₁ = 2.4. At T = 2, p₁ = 1/[1 + exp((2.4−0.8)/2)] ≈ 0.310. At T = 1, it is about 0.168. At T = 0.25, it falls to about 0.00166. The ordering never changes;0 remains locally favoured;but the chance of crossing the energy barrier collapses as the system cools.
Thought experiment 4 · Replace the draw with an argmin
Setup. Start inside a basin where every one-pixel alteration raises energy, although changing a whole block would reach a better configuration. Single change. Replace each Gibbs draw by the lowest-energy local candidate. Result. The configuration never moves. Positive-temperature sampling can accept a temporary loss and construct a route out. Randomness is not decorative uncertainty; it changes which regions of the landscape are reachable.
Annealing trades exploration for concentration
Sampling at a fixed temperature answers a distributional question. MAP restoration asks for a mode. Geman and Geman join the two by lowering temperature as updates proceed. Early in the run, the chain explores many configurations. Later, the Gibbs distribution increasingly concentrates on minimum-energy states. The schedule must be slow enough that the system can escape a deep local basin before uphill moves become effectively impossible.
Theorem B gives a sufficient logarithmic condition. In its site-replacement form, temperature is bounded below by a constant divided by log(1+k), with the constant tied to energy barriers and the number of sites in the authors’ general bound. Under the theorem’s conditions, the distribution of X(k) converges to the uniform distribution over the set of global minima. If there is one minimum, its probability tends to one. With several minima, probability is shared among them.
Two qualifications matter. First, convergence in distribution does not normally mean that one sample path eventually enters a minimum and remains there for all later times. The paper notes that pointwise or almost-sure convergence of the configuration generally cannot be demanded. Second, the theoretical constant is devastating in practice. The authors illustrate that for N = 20,000 sites and a unit energy range, reaching T = 0.5 under the conservative bound could require exp(40,000) site visits.
Their actual experiments cool from roughly T = 4 to T = 0.5 over 300–1,000 full sweeps, using C/log(1+k) with C = 3 or 4 and k counted in sweeps. Those schedules are vastly faster than the general proof requires. The experiments are therefore evidence about a practical heuristic, not an empirical verification of the asymptotic guarantee. This separation between theorem and implementation is one of the most valuable lessons in the paper.
Thought experiment 5 · Same formula, different constant
Setup. Two teams both report T(k)=C/log(1+k). Single change. One uses the conservative barrier-based C; the other uses C=3 after pilot experiments. Result. They share a functional form but not a guarantee. The first may be computationally meaningless; the second may work well without satisfying the proof’s bound. Naming a logarithmic schedule is not enough;units, update clock, constant and stopping rule must be stated.
Three image groups demonstrate modelling choices, not a universal benchmark
The experimental section contains three groups of pictures. Each begins with an original, constructs one or more severely degraded observations, and shows restorations at different points in the cooling process. The implementation is a serial raster-scan Gibbs sampler. No prefiltering or postfiltering is applied. The authors intentionally keep the intensity and line models simple, with only pair cliques in the basic intensity model.
Group 1 starts with a sample from an MRF on a 128 × 128 lattice with five intensities and an eight-neighbour system. Two hundred iterations at T=1 generate the original sample. One degradation adds Gaussian noise with standard deviation 1.5 relative to levels 1 through 5; restorations after 25 and 300 sweeps show early and later annealing. Another degradation combines blur, a square-root nonlinear transformation and multiplicative noise with mean 1 and standard deviation 0.1. Again, the paper shows 25- and 300-iteration results.
Group 2 uses a hand-drawn 64 × 64 image with three grey levels. Gaussian noise with standard deviation 0.7 is added. A 1,000-sweep restoration using only the blob-like intensity process reveals visible flaws. Adding the binary line process and running 1,000 sweeps yields a much more faithful boundary. A second corrupted hand-drawn figure tests the same F–L prior under blur and multiplicative noise. The comparison isolates the modelling contribution of the line layer.
Group 3 asks the line process to find a more general boundary in a 64 × 64 roadside photograph. The image is scaled to two dominant levels, then noise with standard deviation 0.5 is added;roughly a zero-decibel exercise under the paper’s description. The line site now has four states: off plus three directions. Four-line clique energies favour continuations (0.90), then mild turns (1.35), then sharp turns (1.80), while parallel or otherwise disfavoured arrangements can receive 2.70. After 100 and 1,000 sweeps, the sign and arrow are circumscribed by a continuous sequence of inferred line elements.
These displays are historically persuasive but methodologically limited. They are selected qualitative cases, not a blinded comparison across a benchmark with repeated noise draws, prespecified metrics and uncertainty intervals. The paper’s contribution is the coherence of model, algorithm and theorem. Its figures show feasibility and mechanism. They do not establish that one prior or cooling schedule dominates later restoration methods across all natural images.
The paper joined graphical modelling, MCMC and optimisation
Several ideas meet in this article. The MRF–Gibbs equivalence connects conditional specifications to joint energy models. The hierarchical F–L field uses latent variables to express structure while keeping local factors. The Gibbs sampler turns full conditionals into a general simulation mechanism. Simulated annealing turns a temperature-indexed family into an optimisation strategy. Image restoration supplies a demanding application in which each part earns its place.
The terminology also travelled. The authors explicitly call their site-replacement method the “Gibbs Sampler”. Later Bayesian computation generalised the pattern far beyond images: repeatedly draw blocks from full conditional distributions when the joint posterior is difficult to sample directly. Gelfand and Smith’s 1990 account helped bring this logic into mainstream Bayesian statistics. In spatial statistics, imaging, graphical models and machine learning, local conditional updates remain a standard reference point even when faster algorithms replace the original raster scan.
Later work also sharpens the limitations. Besag’s iterated conditional modes replaces random draws with local maximisation and is much faster, but sacrifices the global annealing guarantee and can be initialisation-sensitive. Hajek’s cooling analysis characterises logarithmic schedules through the depth of energy barriers, clarifying why the constant matters. Modern MCMC diagnostics emphasise autocorrelation, effective sample size and convergence from multiple chains; a theorem about an infinite-time law is not a diagnostic certificate for a finite restoration run.
MAP itself deserves scrutiny. A joint mode can be unrepresentative in high dimensions, can change under reparameterisation, and discards posterior uncertainty. Pixelwise posterior means may blur boundaries; marginal modes can assemble a configuration with low joint probability; samples preserve uncertainty but may be expensive. The appropriate estimator depends on the loss attached to restoration errors. The 1984 paper acknowledges posterior expectations as feasible through ergodicity, but its displayed results centre on MAP.
Model criticism is equally important. A smoothness prior can hallucinate continuity, erase rare structures or sharpen a boundary that the data barely support. A directional line process can favour the geometry it was designed to see. If the likelihood misstates blur or noise, the prior may compensate and produce a visually pleasing but physically wrong image. In scientific and intelligence settings, such reconstructions should be treated as conditional inferences, accompanied by sensitivity analyses across plausible priors and forward models.
Evidence boundary
The paper proves asymptotic properties for finite positive fields under explicit visiting and cooling conditions. It demonstrates a small set of restoration cases with chosen energy parameters and finite heuristic schedules. It does not prove finite-time mixing, automatic prior adequacy, calibrated uncertainty for the displayed images, superiority over all filters, or recovery of the true scene whenever the recording is ambiguous.
Place in the canon
Paper 0021 follows Hastings’s general proposal correction and the early causal-inference papers because it shows another way a global probabilistic object can be made operational through local conditional structure. Metropolis–Hastings asks how to accept proposals so a chain targets a distribution. Geman and Geman specialise the move: redraw one component exactly from its full conditional, exploit a graph to compute that conditional locally, and vary temperature to turn sampling into optimisation.
For intelligence and agency, the paper supplies a concrete architecture for distributed constraint satisfaction. Each unit acts from local information; no unit sees the whole configuration; yet the collective dynamics represent a coherent global probability law. This is not consciousness and should not be sold as a theory of mind. It is a disciplined example of how local update rules, latent structure and an objective landscape can generate organised global behaviour.
For later machine learning, the paper is a bridge to energy-based models, latent graphical models, MCMC training and probabilistic vision. It also offers a standing warning: an energy function is simultaneously an inference engine and a value system over configurations. When a modern system “restores”, “completes” or “denoises” evidence, its preferred outputs reflect the regularities its model rewards.
Binary Gibbs-restoration laboratory
The laboratory uses an 8 × 8 synthetic scene, a noisy binary observation and a deterministic pseudo-random Gibbs sampler. Data energy penalises disagreement with the observation. Intensity energy penalises unequal neighbours. In the line-aware model, bonds across the known vertical boundary are broken; this is an oracle demonstration of the mechanism, not an edge detector. Choose a scene, prior and temperature, then run 30 sweeps from the observed grid.
Read the result in layers. Observation error counts the flipped measurements before restoration. Final truth error is available because the scene is synthetic. Data mismatches show how far the reconstruction departs from recorded pixels. Intact-bond disagreements show the smoothness cost the prior still sees. A lower total energy means only a better fit under the selected model; truth error can rise when that model is wrong, especially for the checkerboard stress test.
Memory studio
One sentence
A Gibbs image model converts local structural preferences into a global probability law, and local conditional draws can explore or optimise that law.
Three objects
F intensity field.
L latent line field.
G degraded observation.
Three theorems
A relaxation to π at fixed T.
B slow cooling towards global minima.
C ergodic time averages at fixed T.
Three traps
Local dependence is not short-range marginal independence. MAP is not posterior uncertainty. A logarithmic form without its constant is not the theorem.
The update
Choose site → score each value locally → exponentiate at T → normalise → draw → repeat.
Recall ladder
30 seconds local energy, local draw, global field.
5 minutes add line variables and posterior energy.
1 hour separate fixed-T sampling, annealed MAP and finite-time diagnostics.
What remains after the image is restored
The 1984 paper makes an inverse problem conceptually legible. The likelihood describes the recording apparatus. The prior describes a class of scenes. A graph states which variables interact directly. Clique potentials price local configurations. A sampler turns those prices into local random actions. Temperature determines whether the actions explore the distribution or concentrate near its modes. Every layer has a job, and every layer can be challenged.
Its strongest intellectual achievement is the dual use of locality. Local conditional dependence is a modelling statement about an image, but it is also an algorithmic interface. The same neighbours that define what counts as a plausible region determine what a processor must read before refreshing one site. The global partition function remains inaccessible, yet local probability ratios remain computable.
The line process then shows why hidden variables matter. An intensity-only field must choose between cleaning noise and preserving discontinuities. A latent contour layer changes which disagreements count as errors. This is more than tuning a penalty; it is a change in ontology. The model is permitted to say that two adjacent pixels differ because a boundary exists between them.
Annealing adds the most seductive and the most easily overstated part. A sufficiently slow schedule offers a global asymptotic statement, but its safe constant can be unusable. The practical schedule is an experimental choice. Good finite results require monitoring energy, reconstruction quality, sensitivity to initialisation and sensitivity to the prior. The theorem explains a mechanism; it does not remove engineering judgement.
For contemporary readers, the final lesson is ethical as well as technical. A restored image can look more authoritative than the evidence that produced it. Bayesian coherence does not turn prior-favoured detail into observed fact. Whenever a reconstruction informs a scientific, medical or intelligence decision, retain the degraded input, publish the forward model, compare plausible priors, display uncertainty and mark which features are supported by data rather than supplied by regularisation.