Emotion Recognition from Partially Occluded Facial Images using Prototypical Networks

Facial expression recognition is a challenging task due to the variations and spontaneity of expressions in the real world scenarios. These variations include different head poses, occlusion of face regions, illumination changes and other noises in the image during capturing or transmission. This work aims at classifying five basic human emotions from partially occluded face images. Lack of base line datasets with occluded images is a hurdle to build a model which can generalize well. Meta-learning algorithms offer a solution to this problem. We have carried out research on leveraging meta-learning concepts with prototypical networks in our work to classify emotions in few-shot regime. We have used CMU Multi-PIE database, AffectNet database and images collected from Internet for training and testing purposes. The proposed method is named as MERO (meta-learning for emotion recognition under occlusion).