Robust Multi-view Subspace Learning Through Structured Low-Rank Matrix Recovery

Multi-view data exists widely in our daily life. A popular approach to deal with multi-view data is the multi-view subspace learning (MvSL), which projects multi-view data into a common latent subspace to learn more powerful representation. Low-rank representation (LRR) in recent years has been adopted to design MvSL methods. Despite promising results obtained on real applications, existing methods are incapable of handling the scenario when large view divergence exists among multi-view data. To tackle this problem, we propose a novel framework based on structured low-rank matrix recovery. Specifically, we get rid of the framework of graph embedding and introduce class-label matrix to flexibly design a supervised low-rank model, which successfully learns a discriminative common subspace and discovers the invariant features shared by multi-view data. Experiments conducted on CMU PIE show that the proposed method achieves the state-of-the-art performance. Performance comparison under different random noise disturbance is also given to illustrate the robustness of our model.