An automated system for facial expression recognition is always desirable. However, it is a challenging issue due to the richness and ambiguities with daily facial expressions. This paper presents an efficient approach to recognition of facial expressions of interest. By integrating dynamic Bayesian network (DBN) with the general facial expression language (FACS), a task-oriented stochastic and temporal framework is constructed to systematically represent and recognize facial expressions. Based on the DBN, analysis results from previous periods and prior knowledge of the application domain can be integrated both spatially and temporally. With the top-down inference, the system can make dynamic and active selection among multiple sensing channels so as to achieve efficient recognition. With the bottom-up inference from observed evidences, the current facial expression can be classified with a desired confident level via belief propagation. We apply this task-oriented framework to fatigue facial expression analysis. Experimental results verify the high efficiency of our approach.
Sixth IEEE International Conference on Automatic Face and Gesture Recognition, 2004.