We present a facial recognition system based on a probabilistic approach to adaptive processing of Human Face Tree Structures. Human Face Tree Structures are made up of holistic and localized Gabor Features. We propose extending the recursive neural network model by Frasconi et. al.  in which its learning algorithm was carried out by the conventional supervised back propagation learning through the tree structures, by making use of probabilistic estimates to acquire discrimination and obtain smooth discriminant boundaries at the structural pattern recognition. Our proposed learning framework of this probabilistic structured model is hybrid learning in locally unsupervised for parameters in mixture models and in globally supervised for weights in feed-forward models. The capabilities of the model in a facial recognition system are evaluated. The experimental results demonstrate that the proposed model significantly improved the recognition rate in terms of generalization.