Probabilistic inference of facial action unit by Dynamic Bayesian Network for image sequence

Facial expressions play important role in many applications especially in video surveillance and face identification. The movements of the facial muscles convey the emotional state of an individual. However, reliable expression recognition by machine is still a challenge. This paper proposes a probabilistic framework by Dynamic Bayesian Network (DBN) to study the relationship among Action Units (AUs) that are involved in expressions appearing in an image sequence from video. The image sequence involves the evolution of expressions from neutral to apex and back to neutral which gives very important information to study the temporal changes of AUs. The image sequence as training data is used to learn parameters for DBN which gives transition and observation model of the network. A set of test data is used to be input to the network as evidence to infer the presence or absence of the AU in the image. Experiments are carried out to show the performance of the network which are modeled to represent the true relationship among the AUs.