<p>Facial emotion recognition technology is essential in fields such as intelligent interaction and mental health assessment, yet it faces significant challenges in real-world, unconstrained scenarios: variations in illumination cause feature extraction bias, pose changes and occlusions lead to the loss of key information, low-resolution images hinder the capture of subtle expressions, and label noise together with cross-domain differences further reduce model stability, severely limiting practical applications. To address these issues, this study proposes a novel facial expression recognition framework that enhances robustness through a reliability-aware hierarchical region perturbation strategy and consistency supervision, enabling the model to adaptively switch among fine-grained local perturbation, coarse face-layout-guided perturbation, and global random erasing under an image-quality-aware hierarchical routing mechanism. In addition, the proposed method employs multi-task joint learning that integrates discrete expression classification with continuous valence and arousal prediction, supported by a multi-task loss to promote cross-task knowledge transfer, and incorporates cross-layer attention together with multi-scale feature fusion based on deep residual networks to strengthen the extraction of key-region features and multi-level semantic associations. Experimental results show that the proposed model achieves 69.21% accuracy on AffectNet, 95.12% on RAF-DB, and 80.34% on FER2013, while maintaining a balance between parameter size and computational cost. The proposed framework provides an effective FER solution for complex scenarios and offers broad application prospects in intelligent human–computer interaction and related fields.</p>