Robust sparse coding for subspace learning

Conventional sparse coding fails to learn a robust subspace when the data is corrupted by noises or outliers. To remedy this problem, matrix completion is considered into sparse coding to recover the corrupted data by the normal data and the robust sparse representations can be learned from the recovered data and the normal data. Therefore, a robust sparse coding method, called RSC, is proposed to learn a low-dimensional subspace from the corrupted data. Experiments are carried out on the image dataset which is contaminated by noises or outliers. It is demonstrated that our proposed RSC is more effective and robust in subspace learning and image clustering than other dimensionality reduction methods.