Multiscale Facial Expression Recognition Based on Dynamic Global and Static Local Attention

To better characterize the differences in category features in Facial Expression Recognition (FER) tasks, and improve inter-class separability and intra-class compactness, we propose a Multiscale Facial Expression Recognition model based on dynamic global and static local attention (MFER) from the perspectives of intra-class and inter-class features. Firstly, we propose Dynamic global and Static local attention (DS Attention) mechanism that fuse contextual information, learn potential regions of global and local features between different expression categories, and represent feature discrepancies between categories to distinguish between different expression categories. Then, we design a Deep Smooth Feature loss function (DSF) to balance the probability difference of encoded intra-class features and promote intra-class features towards corresponding centers. Finally, we construct a Multiscale classifier method (Msc) to learn high-frequency and low-frequency information in the dimensional space, represent deep features of multiscale dimensional space, and alleviate sparse distribution problems in high-dimensional space. Experimental results on public datasets RAF-DB, AffectNet-7, AffectNet-8, and FERPlus show that the proposed model achieves state-of-the-art performance with recognition accuracies of 92.08%, 67.06%, 63.15%, and 91.09%, respectively. Code available at https://github.com/XuJ1E/MFER/.