Image Restoration with Locally Selected Class-Adapted Models

In this paper, we propose an automatic way to select class-adapted Gaussian mixture priors, in image denoising or deblurring tasks. We follow the Bayesian perspective and use a maximum a posteriori criterion to determine the model that best explains each observed patch. In situations where it is not possible to learn a model from the observed image, e.g. blurred images, this approach, in general, leads to better results in comparison with using only a generic model. In particular, when the input image belongs to a class for which we have a pre-computed model, the improvement is more pronounced.