Facial Expression Recognition Using Segmentation Residual Network

With the development of artificial intelligence and computer vision, facial expression recognition (FER) plays an increasingly important role in emotional communication. ResNet is a classical convolutional neural network that can be used for expression recognition. However, the performance of the model on FER2013 is not very good, because its ability to extract effective features is insufficient, and the key to the FER task is to extract effective features that help facial expression recognition. This paper proposes a Segmentation Residual Network (SRNet) network based on ResNet, which inherits ResNet ‘s residual module and ResNet ‘s similar architecture. The key of the network is to replace the Residual Block of ResNet with the Segmentation Residual Block (SRB) module designed in this paper. SRB is designed based on ResNet ‘s Residual Block and Attention U-net segmentation blocks. The network also integrates Spatial Transformer Network (STN) and Squeeze-and-Excitation Networks (SENet). Compared with the traditional ResNet, SRNet can extract effective features that are helpful for facial expression recognition and is more sensitive to the spatial transformation of facial features. The experimental results show that the SRNet model outperforms the state-of-the-art deep learning models in terms of sentiment classification accuracy on the FER2013 dataset.