Influence of different feature selection approaches on the performance of emotion recognition methods based on SVM

Abstract

In this paper we evaluate performance of modern emotion recognition methods. Our task is to classify emotions as basic 8 categories: anger, contempt, disgust, fear, happy, sadness, surprise and neutral. CK+ dataset is used in all experiments. We apply Adaptive Boosting and Principal Component Analysis for dimensionality reduction and Support Vector Machine for classification. Size of train dataset is increased by use of few frames of sequences instead of one and vertical mirroring of faces. All images were normalized with mean centering and standardizing. In total 4428 images were used in experiment. The proposed method can work in real time and achieved average accuracy higher than 95%.