Video-based person-dependent and person-independent facial emotion recognition

Abstract

Facial emotion recognition is a challenging problem that has attracted the attention of researchers in the last decade. In this paper, we present a system for facial emotion recognition in video sequences. Then, we evaluate the system for a person-dependent and person-independent cases. Depending on the purpose of the designed system, the importance of training a personalized model versus a non-personalized one differs. In this paper, first, we compute 60 geometric features for video frames of two datasets, namely RML and SAVEE databases. In the next step, k-means clustering is applied to the geometric features to select k most discriminant frames for each video clip. Then, we employ various classifiers like linear support vector machine (SVM) and Gaussian SVM to find the best representative k. Finally, five pre-trained convolutional neural networks, namely VGG-16, VGG-19, ResNet-50, AlexNet, and GoogleNet, were used evaluating two scenarios: person-dependent and person-independent emotion recognition. Additionally, the effect of geometric features in keyframe selection for a person-dependent and person-independent scenarios is studied based on different regions of the face. Also, the extracted features by CNNs are visualized using the t-distributed stochastic neighbor embedding algorithm to study the discriminative ability in these scenarios. Experiments show that person-dependent systems result in higher accuracy and suitable to be used in personalized systems.