Research On Large-scale Face Clustering Based On Graph Structure And Relationship Learning

Large-scale face clustering technology automatically groups face images based on potential identity information,and provides high-quality pseudo-labels for face recognition models,which alleviates the time and labor costs.Recently,with the development of neural network,researchers try to introduce single-sample densities,paired-sample descriptions,and subgraph relationship learning to infer effective clustering patterns to improve clustering results.However,due to the large-scale and complex distribution,there still exist some problems such as high computation time and susceptibility to noisy data.To solve these problems,this paper proposes two advanced face clustering models from relational learning and graph structure learning perspectives.The main study of this paper is as follows:(1)To address the problems of computational inefficiency and sensitivity to postprocessing thresholds of current large-scale face clustering algorithms,this paper proposes the Structure-Enhanced Pairwise Feature Learning(SEPFL)face clustering algorithm,which considers each pairwise relationship between two samples as a learning unit and infers clustering assignments by evaluating a group of pairwise connections.Specifically,this paper determines sample pair connectivity relationships by connection prediction instead of thresholds and adaptively generates pairwise features with high discriminative power using mixed neighborhood information for subsequent clustering identification.Moreover,this paper designs a combined density-guided pair selection strategy to select representative pairs,thus ensuring training effectiveness and inference efficiency.(2)Since Graph Convolutional Networks(GCNs)are susceptible to noisy edges in the graph during face clustering,this paper proposes the Progressive Structure Enhancement Graph Convolutional Network(PSE-GCN)for face clustering.The PSE-GCN combines graph structure learning with feature learning and proposes the Dynamic Graph Construction(DGC)module,which improves the quality of a dynamic graph by enhancing local relationships and suppressing global noise.In addition,to alleviate that GCN is difficult to process large-scale graph data,this work divides the original data into multiple subgraphs and designs a Subgraph-based Neighborhood Re-ranking(SNR)mechanism,which effectively improves the processing efficiency and clustering performance by introducing structural similarity in building subgraphs to reduce the number of noisy samples within the subgraphs.Experimental results on multiple large-scale datasets show that:(1)The SEPFL can balance performance and efficiency while achieving higher accuracy and higher computing efficiency,and without post-processing thresholds,which effectively solves the dependence on post-processing thresholds.(2)The PSE-GCN can effectively mitigate the influence of noisy edges within the graph structure on the clustering results,and achieves better performance than other face clustering algorithms,especially in processing large-scale data with stronger adaptability.(3)The SEPFL and PSE-GCN algorithms are also strongly applicable to non-face data.These research results provide strong support for face clustering and also provide methodological references and experimental bases for the application of clustering methods in data analysis.