I revisit the classical inverse problem of reconstructing a continuous restored image from a small set of blurred and noisy measurements. A Bayesian estimation framework using mixtures of local linear regressors addresses the strong numerical ill‑conditioning that arises when estimating many collaborative local models from few samples. The method models the classical blur‑and‑sampling data formation and performs maximum a posteriori (MAP) imputation to regularize estimation under missing data. I also propose a progressive model selection strategy, where components are adaptively split during an expectation–maximization procedure so that the mixture complexity follows the local spatial structure of the latent image. This yields a compact, continuous, and in‑core representation of the intensity field from which raster images of arbitrary definition can be generated on demand.