Face clustering based on graph neural network for millions of data

Face clustering is an important tool for utilizing unlabeled face data and has a wide range of applications in face annotation and retrieval, etc. How to cluster effectively, especially on large-scale (e.g. millions or more) datasets, is an open problem. Recent studies have shown that graph convolutional neural network (GCN)-based clustering can significantly improve performance. However, these methods require the generation of a large number of overlapping subgraphs, which severely limits the accuracy and efficiency of the model. Since these GCN algorithms have not analyzed the impact of different data features on the models, they usually show excellent performance only on specific datasets. In this paper, a DBSCAN-based graph convolutional network model is proposedby comprehensively analyzing the effects of differences in the distribution of distances and the number of instances on the model. By forming a quadratic clustering model through two distances, the dependence of DBSCAN on distance is eliminated, and the model accuracy is improved by up to 20% in multiple datasets.The robustness of the model is further improved by exploring the fusion of one-hot feature encoding method and multiple adjacency matrix composition methods. The problem of dynamic changes in population size is solved by the adjacency matrix sparsification algorithm. Experiments on several large benchmarks show that our algorithm improves accuracy by 2%~7% compared with existing GCN algorithms, and reduces hardware requirements and improves operational efficiency, which can be applied to million-level face clustering scenarios.