A statistical learning approach to color demosaicing

A statistical learning/inference framework for color demosaicing is presented. Realizing that color is constant over large patches, we regard it as a classical parameter in the loading matrix of a linear observation model, with brightness taking on the role of a latent variable. The former is learned in an unsupervised manner, the latter then inferred from the data. Our framework readily accepts learned prior knowledge as a plug-in regression function of the local strength of color alignment on the background's color and illumination.