Facial Expression Recognition with Squeeze-and-Excitation Network

Abstract

In this paper, we aim to propose a facial expression recognition algorithm by utilizing the spatial attention mechanism. We adopt 3 basic CNN models, including AlexNet, VGGNet and ResNet for comparison. Then we add the attention module, i.e., SENet to VGGNet and ResNet for futher feature enhancement. Our results on FER2013 show the effectiveness of our method. Our VGG+SENet achieves 65.0% accuracy and our ResNet+SENet achieves 66.8% accuracy. Both method with attention can get an obvious performance promotion, which validates the effectiveness of attention mechanism.