Uncertainty-Aware and Class-Balanced Facial Expression Recognition

Facial expression recognition (FER) is a challenging task because of the uncertainty and data imbalance in the real wild. In order to address the above-mentioned challenges, an uncertainty-aware and class-balanced model for FER is proposed. Firstly, the label distribution generation network is designed to represent the characteristics of facial expression distribution comprehensively. The importance of weighting loss based on label distribution is proposed to suppress uncertainty for the model by utilizing the important values to reweight the output results. Secondly, the distance constraint loss is exploited to expand the distance value of the average importance values between groups of different importance values. The relabeling module detects and corrects noisy labels caused by uncertainty based on the sample category prediction vector. Finally, class-balanced loss based on label distribution is proposed to address data imbalance. The learning of difficult samples is enhanced by assigning the weight of samples according to the number of difficult and easy samples. The experimental results show that the proposed model achieves 88.36% accuracy on the RAF dataset. Its performance outperformed the most of FER models and proved the effectiveness of the proposed model.