Facial Expression Recognition using Convolutional Neural Network on Graphs

Since convolutional neural network on graph (GCN) can process data with non-Euclidean structure compared with convolutional neural network, this paper constructs GCN network as a classifier of facial expression recognition and proposes a novel method of combining fixed points with random points to construct undirected graph from the face image. Firstly, face image is transformed into undirected graph by the method of combining fixed and random points. Then, we put the undirected graph into trained GCN and get the result of facial expression. The GCN constructed in this paper consists of six graph convolution layers, a fully connection layer and a SoftMax layer. Each graph convolution layer includes a signal filtering layer and a graph coarsening layer. The proposed method is evaluated on CK + and JAFFE datasets. The results show that the recognition rate of facial expression based on the method of combining fixed points with random points is higher than that based on the method of only using fixed points, and higher than that of using convolutional neural network.