Label Smoothing Facial Expression Recognition Method Based on Resnet

In view of the issues that traditional facial expression recognition models exhibit significant differences in feature distribution for the same face under varying lighting and poses, as well as uneven channel information fusion, this paper proposes a label-smoothing facial expression recognition method. This method is based on the ResNet network, introduces the SE attention module to capture multi-scale features, and uses mixup to enhance unbalanced facial expression data. At the same time, the smoothed cross entropy loss function replaces the traditional cross entropy loss function to further solve the data imbalance problem. Experimental results show that the model achieves an accuracy of 73.69% on the FER2013 dataset, verifying its effectiveness and superiority.