Sparse representation (SR) has played a critical role in the field of biometric image recognition for its strong robustness to noise interference. An important condition for making SR-based methods work is that the training data are totally uncorrupted. However, it may not be guaranteed in practical applications since the image acquisition process could not be carried out in a completely ideal circumstance. This paper aims to address this challenging issue by performing adaptive dictionary learning in the training stage. Following the ℓp norm and adaptive weighted Schatten-p regularization, the learned dictionary is decomposed into a low-rank component, a sparse component, and a dense noise component, where the low-rank one stands for the intrinsic class-specific structure textures, the sparse one stands for the sparse intra-class salient variations, and the noise one stands for the interference information in the corrupted training data, respectively. Furthermore, to facilitate the proposed dictionary learning framework, an improved superposed SR-based classification method is proposed by employing the ℓp norm regularization, which achieves a sparser representation compared with the ℓ1 norm regularization. This effectively enhances the representation precision since the noise interference is separately approximately represented by the decomposed dictionary. Finally, the proposed method is applied in the field of biometric image recognition, such as the palmprint recognition and the face recognition, to validate its effectiveness and improvement. Extensive experimental results on benchmark datasets show that the proposed method can effectively enhance the noise robustness and improve the recognition accuracy compared with some state-of-the-art methods.