Hierarchical multistream recognition of facial expressions

Achieving optimal recognition accuracy, particularly under conditions of input data variability, is a challenge for automatic facial expression recognition. However, little research has been devoted to investigating the robustness of automatic expression recognition under adverse conditions. A facial expression modelling approach is proposed for enhancing the robustness of expression recognition. The approach is founded on hierarchical state-based modelling of streams that represent spatially localised expression dynamics. Experimental assessment shows that the proposed model achieves high and stable recognition accuracy over a range of input data degradation. Moreover, interstream coupling as well as the inclusion of adaptive estimation of model reliability and credibility are shown to make a positive contribution to recognition accuracy.