Answers are grounded in this paper and cite their sources.
Facial expressions are an important means of conveying human emotions. Accurately distinguishing similar expressions requires capturing subtle appearance differences across facial regions, while a single class label cannot fully represent the continuous affective relationships among expression categories. To this end, we propose a Valence-Arousal-Guided Region Attention Network (VA-RANet) for joint discrete expression classification and continuous valence-arousal (VA) estimation. The model learns multiple spatial weight maps from shared features, constructs regional representations through weighted aggregation, and fuses them with global features. A region complementarity constraint is employed to limit the similarity among the spatial weight maps and encourage differentiated regional responses. During training, the distances between the samples and the fixed class prototypes in the VA space are converted into a probability distribution and subsequently fused with the discrete expression labels to generate soft labels that provide auxiliary supervision for discrete expression classification. Ground-truth VA annotations are used for AffectNet, whereas VA pseudo-labels obtained through transfer are used for RAF-DB and CAER-S. Only facial images are required at inference. The model achieves classification accuracies of 65.71% and 62.93% under the seven-class and eight-class AffectNet settings, respectively. Experimental results indicate that the combination of differentiated multi-region visual representations and VA-guided supervision contributes to improved facial expression recognition performance.
No takes yet. Share an insight, caveat, or question.
Xue et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: