Currently, the two-dimensional principal component analysis (2DPCA) method has been widely applied in facial recognition due to its inherent low-covariance matrix characteristic. However, in the real world, captured facial images are often subject to noise interference, leading to issues such as unstable feature extraction and the inability to effectively perceive key discriminative information between categories in many existing 2DPCA-based methods. To overcome these limitations, a novel robust discriminative 2DPCA (RD-2DPCA) method is proposed and applied to facial recognition. First, this method introduces a prototype-driven momentum contrastive learning mechanism. Through momentum updates and online clustering, it constructs semantically consistent positive-negative sample pairs to reduce intra-class divergence, thereby forming a more separable discriminative subspace projection. Second, an adaptive sine angle constraint is embedded in the projection. This constraint ensures a strong association between the solution and the weighted covariance matrix that preserves the global data structure, while further mitigating the influence of outliers on the final projection direction. Finally, to address the optimization problem, a combined strategy of Riemannian gradient updates and a custom non-greedy iterative algorithm ensures stable convergence to a robust optimal solution guided by discriminative principles. To validate the feasibility of this method, experiments were conducted on three publicly available facial datasets, achieving optimal recognition accuracies of 73.90%, 61.31%, and 87.74% respectively. Comprehensive experimental results demonstrate that RD-2DPCA exhibits superior performance compared to single-stage or pure 2DPCA variants in terms of discriminative capability, robustness, and convergence stability. This fully validates the method’s interference resistance and practical application value in complex observation environments.