Traditional robust non-negative matrix factorization methods cannot achieve robust low-dimensional features while the data is heavily contaminated by Salt and Pepper noise.To address this issue,this paper proposes a more robust weighted Manhattan non-negative matrix factorization to recover the contaminated data and obtain robust part-based representations.Our proposed model can be formulated as a non-convex and non-smooth problem,which can be solved by the accelerated gradient method and the rank-one-residual-iteration method.Experiments on the ORL face dataset contaminated by Salt and Pepper noise demonstrate that our proposed algorithm is more effective and robust in image recovery and feature representation learning.