FEAR: Facial Expression Analysis Recognition

FEAR is an automatic facial expression recognition framework combining Active Appearance Model (AAM) and Support Vector Machines (SVM). It classifies seven different expressions, representing the neutral face and the facial emotions of happiness, sadness, surprise, anger, fear and disgust were analysed. The proposed solution starts by describing the human face by an AAM model, projecting the appearance results to the hyperplane that maximizes class separability using a multiclass SVM to emphasize the different expression categories. Results from a Linear Discriminant Analysis (LDA) combined with Mahalanobis distance metric are also given.

Facial Expression Recognition
The facial expression recognition procedure is performed by firstly describing a set of faces using the AAM model. Two classifications aproaches were evalualted, projecting each apperance vector: into Fisherspace and classifying-it using the Mahalanobis distance; and into the hyperspace that maximize class separability using a multiclass one-against-all Support Vector Machines (SVM).

Experimental Results
An expression database was built, consisting in 21 individuals presenting 7 different facial expressions each, namely neutral expression, happiness, sadness, surprise, anger, fear and disgust. The data set is therefore formed by a total of 147 colour images 640x480. Figure 1 shows AAM model instances of a given subject for each one of the 7 facial expressions used.

A leave-one-out cross-validation scheme was used to evaluate the classification. Table 1 shows results of confusion matrices from a LDA combined with a Mahalanobis distance metric and a multiclass SVM.



From the results, is shown that there is an effect of confusion between neutral and sad expressions and also between anger and disgust, suggesting appearance correlation between these two pairs of expressions. Eliminating the confusion effect, faces expressing sadness and anger were excluded.