Feature Selection for Multi-SVM Classifiers in Facial Expression Classification

Facial expressions are non-verbal signs that play an important role between interpersonal communications and also they are the most effective way to describe human emotions. The correct and fast extraction/recognition of the facial expressions is an ongoing research area for computer vision. In this study, the effects of feature selection on the classification of seven different facial expressions (anger, disgust, fear, joy, neutral, sadness, and surprise) are analyzed. To this end, the features for each expression were extracted using Gabor filters from the facial images and selected the best ones using two different feature categories. In the first category, the features that they do exist in one class but do not exist in all other classes have been selected. In the second category, the features that they represent the one class have been selected. These selected features were employed for expression classification using Support Vector Machines (SVMs). Comparative results are given in terms of types for selecting features, feature selection algorithms and the approaches used for the multi-classification. The best results were obtained using RFE algorithm with first type of feature selection and using one-vs-rest approach for training. It was also demonstrated that the feature selection has a positive effect to increase the classification performance comparing when the results were observed with no feature selection.