Facial expression recognition (FER) is a significant challenge due to the complexity of multi-class categorization and the subtle differences between emotional factors, with existing deep learning models often struggling to leverage hierarchical local features for accurate distinction. To find a solution, this study proposes a two-method approach, aiming to enhance FER accuracy by refining classification through the arousal-valence dimension and improving feature extraction via an adaptive attention mechanism to resolve the difficulty of differentiating semantically close expressions. The proposed methodology combines (1) a two-level classification strategy, first, performing fine-grained classification, and second, multi-class categorization by grouping expressions into neutral, positive, and negative classes based on emotional dimensions. (2) a hierarchical attention model that mitigates the challenge of fine-grained expression recognition by stacking global, regional, and local attention in a redesigned spatial hierarchy, first learning a robust semantic grouping (neutral, positive, and negative) of expressions. This approach reduces high-level confusion and provides a clearer pathway for subsequent fine-grained classification. This module extract emotion feature from facial images, all implemented on an enhanced VGG-16 backbone. The experiments on CK+, FER+, AffectNet, and RAF-basic datasets demonstrate that the proposed model achieves superior accuracy and robustness compared to existing methods, effectively capturing multi-level features and leveraging hierarchical subdivision for improved performance. The results confirm that our approach, through its fusion of multi-class grouping and hierarchical spatial attention, offers a more reliable and sophisticated solution for real-world emotion recognition applications.
Oday et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: