<p>Facial expression recognition (FER) in unconstrained scenarios remains difficult due to diverse facial variations and background disturbances. In this work, we propose FRD-BFAN, a feature-region delineation and bilateral feature attention network for robust in-the-wild FER. FRD-BFAN adopts a hierarchical region modeling scheme implemented by a three-branch architecture, where a global branch captures holistic semantics, while local–global and local–local branches progressively refine discriminative regional cues through coarse-to-fine partitioning. To further enhance informative responses and suppress irrelevant patterns, we introduce a bilateral feature attention module that jointly recalibrates features along spatial and channel dimensions, and employ a weighted fusion with multi-branch supervision to stabilize optimization. Experiments on multiple in-the-wild benchmarks demonstrate that FRD-BFAN consistently improves recognition performance and robustness over recent methods, validating the effectiveness of the proposed region delineation and bilateral attention design.</p>