A statistical learning/inference framework for color demosaicing is presented. We start with simplistic assumptions about color constancy, and recast color demosaicing as a blind linear inverse problem: color parameterizes the unknown kernel, while brightness takes on the role of a latent variable. An expectation-maximization algorithm naturally suggests itself for the estimation of them both. Then, as we gradually broaden the family of hypothesis where color is learned, we let our demosaicing behave adaptively, in a manner that reflects our prior knowledge about the statistics of color images. We show that we can incorporate realistic, learned priors without essentially changing the complexity of the simple expectation-maximization algorithm we started with.
View on arXiv