DEEP LEARNING FOR FOR LARGE-SCALE FACIAL EXPRESSION RECOGNITION IN THE WILD

Human facial expression recognition (FER) is a challenging task due to ambiguous facial expressions and highly skewed, imbalanced datasets. Even with recent advances in deep learning, these obstacles still remain. Facial expressions are commonly assigned discrete emotions, as per the categorical model of affect. However, continuous dimensional models of affect, such as those using valence/arousal, provide significant advantages over categorical models in terms of representing complex human emotional states. In this thesis, a novel method to derive dimensional annotations from typical categorical affect datasets will be proposed. The thesis further shows that dimensional and categorical emotion spaces can be jointly learnt via label fusion in order to obtain better feature representations. Experiments on in-the-wild AffectNet as well as a collection of laboratory based FER datasets suggests that jointly learning both dimensional and categorical models of affect significantly reduce redundancy and computational load whilst improving generalization performance, outperforming state-of-the-art methods for dimensional AffectNet with 0.33 and 0.30 RMSEs.

Author dc:creator

• NEO YUAN RONG DEXTER

• Emotions, Deep Learning, Facial Expression Recognition