Hierarchical facial expression recognition

Xiang Zhang, Yu Gu, Huan Yan, Yantong Wang, Fuji Ren, Yusheng Ji

Abstract: Learning facial expressions constitutes a challenging job due to the uncertainties caused by the ambiguity of facial expressions. To address this issue, we propose a simple yet efficient Coarse-to-Fine Network (CFNet) inspired by human being's cognitive mode for suppressing such uncertainties. A child learns quickly whether his behavior is allowed by reading adults' coarse facial expressions like positive or negative, and then adjusts accordingly via further interpreting its fine meaning like happy or angry. Similarly, CFNet first aggregates basic facial expressions into 3 coarse categories based on their distributions in the Valence-Arousal (VA) emotion space. Then, CFNet leverages fine labels with the coarse classification results for fine-grained facial expression recognition with discriminative loss handling high intra-class variations within coarse categories. Experiments on benchmark datasets demonstrate the superiority of our method over the state-of-art rivals.