Facial Expression Recognition Based on Feature Representation Learning and Clustering-Based Attention Mechanism

Facial expression recognition (FER) plays an important role in many computer vision applications. Generally, FER networks are trained based on the annotated labels or the probability distribution that an expression belongs to seven expression categories. However, the quality of annotations is heavily affected by ambiguous and indistinguishable facial expressions caused by compound and mixed emotions. Furthermore, it is difficult to annotate the seven-dimensional labels (probability distributions). To address these problems, this paper proposes a new FER network model. This model represents each type of facial expression as a high dimensional feature vector, based on which the FER network is trained. The high-dimensional feature representation of each facial expression class is learned by a special binary feature representation generator network. We also develop a clustering-based group split attention mechanism, which enhances the emotion-related features effectively. The experimental results on two lab-controlled datasets and four in-the-wild datasets demonstrate the effectiveness of the proposed FER model by showing clear performance improvements over other state-of-the-art FER methods. Codes are available at https://github.com/Gabrella/GLA-FNet.