| With the development of technology and the Internet,face recognition,as one of the popular biometric technologies,has been widely used in fields such as security,entertainment,and education.Thanks to the advancement of deep learning,the performance of face recognition technology has also been significantly improved.However,the high accuracy of face recognition at this stage relies on the training of a large number of annotated data sets on the model.The more annotated data,the better the performance of face recognition.Although a large number of face images can be easily obtained from the Internet,manually labeling them is expensive and time-consuming.In order to solve this problem,people turn their attention to clustering methods.Traditional clustering methods have limitations.For example,k-means requires the clusters to be convex,DBSCAN requires the data to be uniformly distributed,and Spectral Clustering is more sensitive to parameters.It is unable to deal with the complex face structure in real life,so it is difficult to deal with the actual face clustering problem.Recent studies have shown that supervised clustering methods can significantly improve its performance.The face image is regarded as a node,and a subgraph is constructed for each node to capture contextual information and determine whether there are edges between the nodes.These edges can help construct clusters.However,there are overlapping parts between subgraphs,especially the subgraphs generated by nodes located near the cluster center.Repeated reasoning for a large number of edges in the subgraph will increase the amount of calculation and reduce the clustering efficiency.This thesis proposes a face clustering method based on graph convolutional neural network(DDC-GCN),which divides the data into high-density and lowdensity parts.The high density is partly at the center of the cluster and has strong connectivity.Only need to do GCN subgraph link reasoning for the low-density part,thereby reducing the amount of calculation.At the same time,in order to ensure sufficient extraction of context information for low-density parts,an adaptive subgraph construction method is proposed,which constructs an adaptive subgraph for each node,and adopts a link prediction model and a node classification model to infer the subgraphs respectively.This thesis conducts experiments on the IJB-B and MS-Celeb-1M data sets.The experimental results show that the clustering accuracy of the proposed method is improved by 2% compared with L-GCN,and the reasoning speed is increased by 5 times. |
