Facial Expression Recognition and the Application of Supervised Contrastive Learning

Abstract

Facial expression is an essential part of communication in human life. In order to automating the process of facial expression recognition and improve the accuracy, in this paper, I explore the problem of facial expression recognition through the FER-2013 dataset with a new loss function—supervised contrastive learning. The goal is to classify similar features close together with the selection of different anchor points and positive and negative examples under the label of pre-classification. I address this task by Convolutional Neural Network, using both a shallow CNN model of my own as well as deeper models such as ResNet, VGG, and Inception. I fine-tuned these models and compared their performances on the dataset. Then I ensembled them to reach a better performance. As a result, I obtained a final model that reaches 70.24% accuracy on the test set, beating two baselines (human recognition rate and null model accuracy) proposed by previous works.