Blind separation of non-stationary images using Markov models

In recent works, we presented a blind image separation method based on a maximum likelihood approach, where we supposed the sources to be stationary, spatially autocorrelated and following Markov models. To make this method more adapted to real-world images, we here propose to extend it to non-stationary image separation. Two approaches, respectively based on blocking and kernel smoothing, are then used for the estimation of source score functions required for implementing the maximum likelihood approach, in order to allow them to vary within images. The performance of the proposed algorithm, tested on both artificial and real images, is compared to the stationary Markovian approach, and then to some classical blind source separation methods.