TSC-GCN: A Face Clustering Method Based on GCN

Jinmin Xue, Shengdong Qu, Jingxian Li, Yan Chu, Zhengkui Wang

Abstract: For face clustering, the density of each cluster distribution in the feature space is different. Too high similarity pruning will lead to the sparse clustering not being divided and reducing the recall ratio, while too low similarity pruning will lead to the decline of clustering accuracy. We propose the Two-Stage Clustering Method Based on Graph Convolutional Neural Network (TSC-GCN), in which the clustering size are set to measure the sparse degree of clustering and pruning with a low similarity degree. The clustering with sparse distribution is screened out, then the requirements for nodes similarity are improved. At the same time, the number of neighbor nodes is set to prevent the clustering core from deviating from the aggregation of nodes, and the clustering with dense distribution is screened out. The experimental results show that TSC-GCN can give a good consideration of the accuracy and recall ratio both, and achieve better clustering effectiveness than the state-of-the-art methods.