Research On Face Clustering Algorithm Based On Graph Convolutional Network

Face clustering is an essential tool to utilize unlabeled face data,and it has a wide range of applications in face database construction and face image annotation.Generally speaking,the merits and demerits of face clustering are mainly affected by the feature extraction method and the design of clustering algorithm.With the application of deep learning in face recognition,deep face model has become the main face feature extraction method at present.Most existing face clustering algorithms are proposed based on convolutional neural network,and similarity measurement is adopted for clustering rules.However,convolutional neural network is more suitable for processing data in Euclidean space.Therefore,how to effectively learn the structural information between clusters of different classes is a great challenge.Since the advent of graph convolutional network in2017,face clustering algorithm models based on graph convolutional network began to appear.But at present,this kind of algorithm is mostly studied by constructing subgraph.This paper studies the face clustering algorithm based on graph convolutional network.A connection-based clustering algorithm framework is proposed,and the clustering problem is transformed into the connection probability of edges between two vertices on a graph.Among them,the input data is constructed by global graph,which solves the problem of overlapping results caused by large quantum graph and introducing noise reduction.At the same time,the edge properties are manipulated directly,rather than using symmetric adjacency matrices.The main work of this thesis is as follows:(1)A face clustering algorithm based on graph convolutional network is proposed to transform the clustering problem into the prediction problem above the graph.The connection-based clustering algorithm can predict whether there is a connection between two vertices without limiting the distribution of the data set,which makes the model perform well even for the face data set with complex distribution.(2)In order to solve the problem that the input data adopts the method of constructing sub-images and the introduction of the prediction results overlap and the need for manual noise reduction,this paper designed a structural approach of the global figure,among them,each face image as a figure of a vertex,according to the included Angle cosine distance measuring the similarity between human faces,and thus construct edge.The final prediction result of the edge is output by the graph convolutional network model.(3)To improve the clustering effect and enhance the expression of input data graph,edge attributes are introduced on the basis of vertex attribute characteristics,and model design is carried out based on Graph Network Block graph network model(a specific graph convolutional network model),and the message passing method of alternating update of vertex and edge is adopted for processing.The proposed model was verified on CASIA,IJB-B,DPF and YTB datasets,respectively.The experimental results fully demonstrate the effectiveness of the proposed face clustering algorithm,and its F-measure and NMI scores are significantly improved.Compared with the existing face clustering algorithms based on graph convolutional networks,the proposed algorithm can solve the clustering problem of face data sets with complex distribution better,and shows better performance.