Deep face clustering using residual graph convolutional network

Face clustering has important applications in image retrieval and criminal investigation. Face images can be seen as the nodes of a graph and the possibility of links between the nodes will help us find clusters. Graph Convolutional Networks (GCNs) are powerful tools to infer the possibility of linkage between a given node and its neighbors. However, existing face clustering methods use shallow GCNs and have limited learning capabilities. We propose a deep face clustering method using Residual Graph Convolutional Network (RGCN), which contains more hidden layers. For each node, k-Nearest Neighbor (kNN) algorithm is used to construct its sub-graphs. Then we apply the idea of ResNet into GCNs and construct RGCN to learn the possibility of linkage between two nodes. Compared with other popular face clustering approaches, our method is more efficient and has better clustering results in the experiments. In addition, the proposed RGCN clustering approach is able to detect the quantity of clusters automatically and can be extended to large datasets.