Facial Sentiment Classification Based on Resnet-18 Model

Recently, human-computer interaction is become increasingly common in daily life, in which facial expression recognition has a wide application potential. At present, the state-of-the-art accuracy of facial expression classification based on the Kaggle database is only 71.20%, which still indicates some room for improvement. Based on the resnet-18 model, we find that a large number of parameters in the model are prone to over-fitting, which hinders the face recognition accuracy. In this paper, we propose a modified network model based on the resnet18-network, in which we change the original average pooling layer to a global average pooling layer with a double convolution layer. Our experiment result suggests that our model outperformed the state-of-the-art results showing a 1.49% increase in accuracy on the Kaggle database.