Neuro-Fuzzy Analysis of Facial Action Units and Expressions

In this paper an accurate real-time sequence-based system for representation, recognition and analysis of lowintensity facial expressions and FAUs is presented. The feature extraction is done using facial feature point tracking and biased discriminant analysis as an efficient dimension reduction method. A novel classification scheme based on neuro-fuzyy modeling of the FAU intensity is presented. In this method, Takagi-Sugeno type ANFIS is used to extract the fuzzy rules. Each rule applies a linear approximation to estimate the FAU intensity in a specific fuzzy subspace. In comparison with the SVMs, the ability of this method to model a highly non-linear system (function) and the fuzzy natural of the FAUs causes high recognition rate of the low-intensity and combined FAUs, on Cohn-Kanade database in presence of the head motion. Moreover, the decision tree as a top-down hierarchical rule-based classifier is used for classification of the basic facial expressions. Applying this classifier, accurate human-interpretable model for facial expression interpretation by continuous value of FAU intensity can be created. These models would be useful in animation, lie detection and behavioral and cognitive sciences areas. KeywordsFacial Activation Units (FAUs), biased discrimnant analysis, Adaptive-Network-Based Fuzzy Inference System (ANFIS), decision trees, computer vision.