A Mixture Model for Concealment of Lost Blocks in Coded Image and Video

When errors are introduced into a compressed image or video bitstream, blocks of pixels are lost. The decoder attempts to conceal this loss to maximize displayed picture quality. We discuss a few different conditions under which satisfactory interpolation of the missing blocks is possible. For images, these conditions include situations where the image is locally smooth or locally one dimensional. For video, in addition to the conditions for images, satisfactory concealment is also possible when the current frame is locally a translation of the previous frame. In order to obtain a concealment strategy which will perform satisfactorily whenever any of these conditions is present, we model images and videos as mixtures of simple Markov processes, where each of the simple processes represent one possible condition for satisfactory concealment. We then use Bayesian reasoning for performing the optimal interpolation using the model.