Optimal feature selection for 3D facial expression recognition using coarse-to-fine classification

Automatic facial expression recognition for novel individuals from 3D face data is a challenging task in pattern analysis. This paper describes a feature selection process for pose-invariant 3D facial expression recognition. The process provides a lower dimensional subspace representation, which is optimized to improve the classification accuracy, retrieved from geometrical localization of facial feature points to classify facial expressions. Fisher criterion-based approach is adopted to provide a basis for the optimal selection of features. Two-stage probabilistic neural network architecture is employed as a classifier to recognize the facial expressions. In the first stage, which can be regarded as the coarse classification, the facial expressions are classified into one of the three expression groups formed using seven basic facial expressions. In the fine classification stage, final expression is determined by using within group classification. Facial expressions such as Neutral, Anger, Disgust, Fear, Happiness, Sadness, and Surprise are successfully recognized with an average recognition rate of 93.72%.