Differentiating Duchenne from Non-Duchenne Smiles Using Active Appearance Models and the Facial Action Coding System
Affect recognition has been a popular area of study in computer vision and psychology over the past decade. According to the psychologist Dr. Paul Ekman, there are 7 "universal" emotions: happiness, sadness, surprise, fear, anger, disgust, and contempt. Happiness, for a long time, was considered to be any facial expression that contained a smile. Dr. Ekman showed that there are actually a number of different types of smiles, many of which are not linked to positive emotion, such as miserable smiles.
Duchenne smiles, which have been linked to genuine enjoyment and are therefore commonly referred to as genuine smiles, include the contraction of the orbicularis oculi, pars lateralis muscle, also known as the Duchenne marker. Non-Duchenne smiles lack this muscle contraction and are therefore commonly referred to as posed smiles. Many studies have attempted to differentiate Duchenne from non-Duchenne smiles using a variety of feature extraction and classification techniques. Many of these studies have used facial emotion expression images that have been labeled by certified Facial Action Coding System (FACS) coders.
In this study, the technique of using Active Appearance Models (AAMs) for feature extraction, and k-Nearest Neighbor (kNN), Naive Bayes, and Support Vector Machines (SVMs) for classification, to differentiate Duchenne from non-Duchenne smiles is evaluated, using images from the Extended Cohn-Kanade AU-coded Facial Expression Database (CK+) and 2D images from the FACS AU-coded Bosphorus 3D Database (Bosphorus). This methodology has not been explored as of yet for the differentiation of Duchenne from non-Duchenne smiles, to the best of the author's knowledge. Three generalization approaches were evaluated: ideal-case (No generalization), semi- generalized, and fully-generalized. For all three approaches, kNN and Naïve Bayes performed favorably to, and Support Vector Machines outperformed, previously published methodologies from similar studies. This indicates that Active Appearance Models may be a very strong choice for feature selection for affect recognition systems
FULL TEXT (pdf)