Discriminative Least Squares Regression (DLSR) improves classification performance by enhancing inter-class separability through the ϵ-dragging method. However, it overlooks the impact of intra-class structure on the model’s discriminative ability and is sensitive to noise interference. While many DLSR variants have been proposed to preserve intra-class structure modeling and noise robustness, they often fail to achieve a balanced trade-off between these goals and model complexity. To solve these challenges, this paper develops a new model, called Structural Enhanced Robust Discriminative Least Squares Regression (SERDLSR). Specifically, SERDLSR introduces a projection distance regularization to promote the compactness of intra-class samples in the projected space. Additionally, this model integrates local sparsity and local low-rank constraints, the former improves intra-class consistency, while the latter preserves the underlying low-dimensional structure of intra-class samples. By integrating the above three types of constraints, SERDLSR effectively strengthens the intra-class structure of projected feature representations while enhancing robustness against the inherent high-dimensional noise in the dataset. Extensive experiments on face recognition (AR, ORL, CMU PIE, FERET, and Georgia Tech datasets), biometric recognition (PolyU Palmprint dataset), and object recognition (COIL-20 dataset) demonstrate that SERDLSR achieves superior classification performance.