Deep Residual Learning for Facial Emotion Recognition

Abstract

Abstract Facial emotion recognition has been an active research topic with its extensive applications in the field of health care, videoconferencing, security and authentication and many more. Human facial emotion recognition (anger, happiness, sadness, etc.) is still a challenging task, and many approaches have been developed. This thesis work proposed a deep residual learning network to recognize human facial emotions which defines approach to train very deep network. ResNet50 with 50 layers is the base model for this thesis work, and the performance is compared with convolutional neural network (CNN) in terms of accuracy, training time and training loss. Two publicly available datasets CK+ and FER were chosen, and the performance of the network was compared using these datasets. The result of the proposed model shows that the deep residual learning network (ResNet50) performs better than CNN for facial emotion recognition in both FER and CK+ datasets. The overall accuracy of the network was 89.8% in comparison with CNN 62% in CK+ dataset. Similarly for FER dataset, the accuracy was 81.9% compared to CNN 66%.KeywordsFacial emotion recognitionDeep residual learning