Robust facial expression recognition with global-local joint representation learning

As an important part in computer vision, facial expression recognition (FER) has received extensive attention, but it still has lots of challenges in this area. One of the important difficulties is to remain the topological information in the feature extraction operation. In this paper, we propose a novel facial expression recognition method with lite dual channel neural network based on graph convolutional networks (DCNN-GCN). In the proposed method, (1) the topological structure information and texture feature of regions of interest (ROIs) are modeled as graphs and processed with graph convolutional network (GCN) to remain the topological features. (2) The local features of ROIs and global features are extracted with dual channel neural networks, which can improve the performance of features extraction and reduce the complexity of networks. The proposed method is evaluated on CK+, Oulu-CASIA and MMI data sets. Experiment results show that the proposed method can significantly improve the accuracy of facial expression recognition. In addition, the network is much lite and suitable for application.