Facial Expressions Classification based on Broad Learning Network

In this paper we propose a new broad architecture network for facial expressions classification. We use Grad-CAM technique to observe which parts of the image are important for classification. We analyze the changes of the expanded width in the network structure which may cause Grad-CAM to shift their focus on the impact parts of various emotional images. They help to establish association between neuromorphic networks and emotion perception. At the same time, we analyze the similar patterns between Grad-Cam map for image of broad architecture network and the electroencephalogram (EEG) heat-map of brain network in facial expressions classification. In order to show the effectiveness of our broad architecture network we test the proposed method on three motion expression classification benchmark datasets: Extended Cohn-Kanade Dataset (CK+), Facial Expression Recognition (FER), JAFFE. The experimental results show the effectiveness of our broad network when compared with vision transformer (ViT).