Robust discriminative nonnegative dictionary learning for occluded face recognition.

•The occlusions can be estimated adaptively via the CIM-induced measurement without any priors for the occlusions.•The class mean regularization term is proposed to enhance the discriminant ability of the low-dimensional representation.•The ℓ21-norm is adopted to robustly select the discriminative features.•Experimental results on synthetic and real datasets demonstrate the proposed method outperforms other methods.