Face clustering using a novel density peaks clustering algorithm

Face clustering remains a challenging task due to the high intra-class variability and uneven density distributions inherent in real-world face datasets. These characteristics often undermine the performance of conventional clustering algorithms. To address these limitations, this paper introduces a novel density-based clustering method, termed DPC-MK (Density Peak Clustering with Mixed k-Nearest Neighbor strategy). Initially, the reverse nearest neighbors and shared nearest neighbors of each sample are identified based on the k-nearest neighbor (KNN) method, and their counts are quantitatively assessed. Subsequently, the distances between each sample and its k-nearest neighbors are computed to evaluate their respective contributions to the local density. The quantified reverse and shared neighbor counts are then integrated with the distance-based density metric to yield an enhanced local density estimate. Using this refined density, the relative distance between each sample and any other point with higher density is computed. A decision graph is then constructed from the modified local density and relative distance values to identify cluster centers. Finally, non-center points are assigned to clusters by following density gradients toward their nearest higher-density neighbors. The results of the ablation study clearly demonstrate the complementary roles of each component as well as the effectiveness of the method we proposed. The efficacy of DPC-MK is further validated on multiple UCI benchmark datasets and public face clustering datasets. Comparative evaluations against baseline and state-of-the-art algorithms—including K-means, DBSCAN, FCM, DPC, DPC-KNN, DPC-NN, DPC-FWSN, and LPMNN-DPC—demonstrate that DPC-MK achieves superior clustering performance and maintains robustness across diverse clustering scenarios and varying cluster counts, highlighting its strong generalization capability.