Probabilistic graphical models are popular tools in statistical signal processing. The dependencies between the random variables described by such models enable to model a large class of statistical problems. Among probabilistic graphical models, Hidden Markov models and their extensions, Pairwise Markov models, are latent variable models which have found applications in image segmentation. In this paper, we address this problem by introducing a new class of Pairwise Markov models whose parametrization allows the use of (deep) neural networks architectures, for example. We focus on the unsupervised parameters estimation in these general models and we show that the combination of our general framework with (deep) neural architectures outperforms classical Pairwise Markov models for the task of unsupervised image segmentation.