Subject-Independent Multi-Domain Facial Emotion Recognition with Local Salient Features

Despite the rich literature on facial expressions recognition, little works have considered the fusion of features extracted in multiple domains. This paper proposes a multi-domain facial emotion recognition framework. The approach extracts simultaneously local facial features from the spatial domain (geometric features), the frequency domain (local wavelets features) and the spatio-temporal domain (Local Binary Patterns on Three Orthogonal Planes (LBP-TOP)). A feed-forward neural networks and support vector machines (SVM) models were trained on each of the extracted features to assign one of the seven basic affective states. A late fusion was then performed to characterize the final facial expression. Experiments were conducted on SAVEE database. Classification results show the complementary effect of the proposed multi-domain features with an overall accuracy of 99.18 % which outperforms state-of-the-art works.