Self-Cure facial expression recognition based on Coordinate attention and Lsland loss
Lulu He, Yian Zhang, Weihuang Lin, Chenyu Jiang, Xuemei Sun · 2023
Facial expression recognition (FER) is a challenging task. The following two significant problems often exist in real-life FER tasks: first, facial expression images in the wild are uncertain, i.e., occlusion or blurred images; and second, there is little intra-class similarity and high inter-class similarity. In this paper, we propose a Self-Curing network based on Coordinate attention and Island loss(CoIsland-SCN) to address the above two problems. Firstly, the coordinate attention module aggregates the input features in vertical and horizontal directions into two independent direction-aware feature mappings, and the encoded attention contains channel, directional and positional information at the same time, which well solves the facial expression uncertainty problem; secondly, the Self-Cure module performs ranking regularization and relabeling according to the image importance and calculates the cross-entropy loss and the ranking regularization loss; and finally, combines the island loss to adjust the model to solve the problem of poor intra-class tightness and poor inter-class separability. The method proposed in this paper achieves an accuracy of 86.41% on the public dataset RAF-DB, which is very superior compared with recent experiments and proves the effectiveness of the proposed method.