In this paper, we propose a relaxation algorithm based on the game theory for scene labeling problems. Using a Bayesian modeling by Markov random fields, we consider the maximization of the a posteriori probability of labelings. We design a (noncooperative) game which yields an easily parallelizable relaxation algorithm. We prove that all the labelings which maximize the a posteriori probability are Nash equilibrium points of the game, and that all the Nash equilibrium points are local maxima. We also prove that our relaxation algorithm converges to a Nash equilibrium. Experimental results show that the algorithm is very efficient and effective, and that it exhibits very fast convergence.