A Novel Low Rank Spectral Clustering Method for Face Identification

Background: Low rank is a recent significant model to explore the inner subspace structure of samples, which has been successfully used in many pattern recognition tasks. Methods: In this paper, we proposed a novel spectral clustering method to address the face identification problem. There are three main contributions in our paper. Firstly, the sparse coding under a cluster-based learned dictionary is taken as the character sample of each face. Secondly, the collaborative low rank representation is incorporated in the comprehensive optimization framework to construct an effective affinity graph iteratively, which is different from the conventional ones to tackle the graph construction and spectral clustering independently. We revised all the patents relating to the face identification. Thirdly, a numerical algorithm is developed to solve the optimization framework and obtain a stable solution. Results: The experimental results showed the superior performance of the proposed method on recognition ratio. Conclusion: It means that our proposed low rank based identification algorithm outperforms the existed excellent methods.