Facial Expression Recognition with Manifold Learning and Graph Convolutional Network

Facial Expression Recognition (FER) has the ability to detect human affect state. Most of the methods employed for FER task do not really consider the correlation among FER data labels to resolve data annotation and ambiguity problems. Label Distribution Learning (LDL) application to FER considerably address this, but only in the presence of data with distribution labels. Therefore, methods that would recover label distribution from logical labels are required. This work is presenting a graph-based label enhancement approach with manifold learning and Graph Convolutional Network (GCN) for facial expression recognition. The manifold learning approach transforms FER data as a graphical problem, where the data points are considered as nearest neighbours represent graph nodes, with the motive of representing the distances along the edges of the neighbouring graph with the approximate distances along the manifold. This process uses the nearest neighbour graph to learn the geometric structure in FER data, which also learn the possible correlation among the data labels. The graphical convolutional network is employed to incorporate the information provided in the manifold learning and the logical description of the data to classify the nodes of the graph using the information of the nearest neighbours. The experiment conducted on the Binghampton University-3D Facial Expression (BU-3DFE) and the Cohn Kanade extension (CK+) data shows that the model gives promising results.