A combined Markovian and Dirichlet sub-mixture modeling for evidence assignment: Application to image fusion

The estimation of Mass functions is a key issue in evidence theory. In this paper, we propose an algorithmic framework to achieve this task using a statistical modeling of the data. The confidence level of each component in the frame of discernment is represented and quantified using a sub-mixture model, where each data cluster is approximated by a Dirichlet distribution. We discuss and show the interest of using the Dirichlet distribution to model sensors corrupted by non-Gaussian noise. The contextual relationship is integrated within the fusion scheme using Markov fields. In this context, we propose an adaptation of the iterated conditional modes (ICM) algorithm which permits to deal with compound hypotheses as defined by Dempster–Shafer theory. The experiments are conducted, in the context of image segmentation using multiple sensors, on synthetic, radar and optical (SPOT) images.