Switching Pairwise Markov Chains for Non Stationary Textured Images Segmentation

Hidden Markov chains (HMCs) have been extensively used to solve a wide range of problems related to computer vision, signal processing (Cappé, O., et al 2005) or bioinformatics (Koski, T., 2001). Such notoriety is due to their ability to recover the hidden data of interest using the entire observable signal thanks to some Bayesian techniques like MPM and MAP. HMCs have then been generalized to pairwise Markov chains (PMCs), which offer similar processing advantages and superior modeling possibilities. However, when applied to nonstationary data like multi-textures images, both HMCs and PMCs fail to produce tolerable results given the mismatch between the estimated model and the data under concern. The recent triplet Markov chains (TMCs) have offered undeniable means to solve such challenging difficulty through the introduction of a third underlying process that may model, for instance, the switches of the model along the signal. In this paper, we propose a new TMC that incorporates a switching PMC to model non stationary images. To validate our model, experiments are carried out on synthetic and real multitextured images in an unsupervised manner.