RGCFace: Regularized Global Center loss for Deep Facial Expression Recognition
Qiong Liu, Yangcheng Zhou, Wei Liu, Guangrui Li, Chenxu Li, Wenbai Chen · 2022
In the facial expression recognition task, we expect that the intra-class facial expression images could be distributed compactness while the inter-class ones could be distributed separability. However, the current loss functions of Convolutional Neural Network (CNN) cannot balance these two aspects obviously at the same time and they are weak in learning discriminative features. In this paper, a novel Regularized Global Center (RGC) loss is proposed to maximize the decision boundary in angular space directly instead of optimizing the inter/intra class distance, so as to enhance the discriminative ability of the deeply learned features by uniformly distributing the class centers on the feature space. In particular, a Normalization layer and a Scale layer (N-S) are introduced to make different features distributed on one hypersphere. Extensive experiments on three public facial expression datasets (CK+, MMI, and Oulu-CASIA) have proved that the proposed method outper-forms the comparable methods on the recognition accuracy.