Towards a General Vision System Based on Symbol-Relation Grammars and Bayesian Networks