Novel Approaches for Facial Emotion Recognition Applied to Small-scale and Large-scale Databases
Facial emotion recognition became a challenging issue for many researchers in different area such as computer vision and artificial intelligence. The difficulty of emotion recognition is due to the variations in facial expressions from person to another from country to another, different cultures and races, pose and illumination variations. Thus this thesis proposed two reliable methods to deal with facial emotion recognition on for small-scale and large-scale databases. The first approach is based on pyramid histogram of oriented gradient(PHOG) feature descriptor, a new genetic algorithm(GA)-linear discriminant analysis(LDA) based on hinge loss function feature selection, and LDA classifier. The proposed approach has been evaluated on Radboud database. It has achieved similar accuracy as convolution neural network(CNN) by using much fewer features than CNN. The second approach is a variation of the new recently released capsule network(CapsNet). This thesis proposed a new model architecture for coping with facial emotion recognition on a large-scale database. Fer2013 database has been used to evaluate the proposed model, this database is one of the largest facial expression databases. The proposed model has outperformed the baseline in terms of classification accuracy