Facial emotion recognition based on improved ResNet

Abstract

A common and difficult task in computer vision is the identification of facial expressions of emotion. Facial emotion recognition has shown great promise in many areas including healthcare, robotic communication and customer service. However, the variability of human appearance and muscle movement leads to difficulties in facial emotion recognition. Therefore, deeper convolutional neural networks are introduced to recognize facial emotions. The residual network (ResNet) can be built in a deep architecture that can solve the degradation problem when the depth of the network increases. In this paper, an improved ResNet50 is proposed to implement facial emotion recognition. Specifically, the proposed model appends two blocks consisting of fully connected layers to the ResnNet50. The layers are stacked with shortcut connections to solve the degradation problem and to make the training process smoother. The accuracy achieved by the improved model on the Facial Emotion Recognition 2013 dataset (FER-2013) is 13.31% higher than that of the ResNet50. Experimental data indicate that the improved model performs efficiently in facial emotion recognition due to shortcut connectivity and the addition of fully connected layers. Meanwhile, the degradation and gradient disappearance problems are improved.