Probabilistic graphical models such as Hidden Markov models have found many applications in signal processing. In this paper, we focus on a particular extension of these models, the Pairwise Markov models. We propose a general parametrization of the probability distributions describing the Pairwise Markov models which enables us to combine them with recent architectures from machine learning such as deep neural networks. In order to evaluate the power of these combined architectures, we focus on the unsupervised image segmentation problem which is particularly challenging and we propose a new parameter estimation algorithm. We show that our models with their associated estimation algorithm outperforms the classical probabilistic models for the task of unsupervised image segmentation.