Face Clustering Algorithm Based on Domain Structural Similarity

A KNN-Struct-ChineseWhispers algorithm based on K-nearest neighbour similarity aggregation is proposed for face clustering to address the problem that the ChineseWhispers algorithm is slow for large-scale face clustering and noisy edges lead to reduced clustering accuracy. Using Faiss to speed up the K-nearest neighbour graph initialisation process, the problem of calculating the similarity between two nodes is transformed into a limited radius K neighbour query for a single node corresponding to the indexed clustering centre. The noisy edges in the graph are filtered by aggregating feature similarity through the neighbourhood structure, and an isolated node detection module is added to improve computational efficiency and clustering accuracy. Based on the CNN+ArcFace Loss method for extracting face features, experiments on several benchmark datasets show that the proposed algorithm can effectively improve the time efficiency and accuracy of clustering compared to existing face clustering algorithms.