Face clustering based on graph embedded semi-NMF with Kernels

Non-negative Matrix Factorization (NMF) has been widely studied and applied to variant computer vision tasks, such as image clustering and pattern classification. Meanwhile, real world stimuli for human neural system (e.g., face images) are usually represented as high-dimensional data vectors rely on graph embedding in original Euclidean space. Thus, the traditional NMF and its variants exhibit weakness for such clustering tasks. In this paper, we propose a Graph Embedding based Semi-NMF on Reproducing Kernel Hillbert Space (RKHS-SNMF) for human face clustering. This algorithm jointly considers the efficient dimension reduction and effective locally geometric preservation based on the exceptional property of measurability and differentiability of Reproducing Kernel Hillbert Space (RKHS). Extensive experimental results show that our RKHS-SNMF clustering algorithm outperforms the traditional clustering methods and the state-of-the-art NMF based clustering approaches on four benchmark human face databases.