Face Mamba: A Facial Emotion Analysis Network Based on VMamba*

In recent years, convolutional neural networks and Transformer and its variants, such as Vision Transformer and Swin Transformer, have been the main models for facial expression recognition tasks. However, both have corresponding problems. In convolutional neural networks (such as CNN), CNN mainly focuses on local features in image processing, has poor grasp of global information, and is difficult to handle large data sets. In Transformer, its performance is limited by factors such as the large number of parameters and the high complexity of attention calculation. Inspired by the Mamba model's ability to model long sequences, we propose a facial emotion analysis network based on VMamba (visual state space model). This network not only reduces the computational complexity to linear by using the selective SSM module and Cross-Scan Module (CSM) feature scanning mechanism in the VMamba model, but also shows good performance in feature extraction. In order to enhance the performance in facial emotion analysis, we also propose the FaceR module to extract features from the input face image. A large number of experiments have proved that the model has good performance in facial expression recognition tasks, with accuracies of 66.57%, 62.65% and 90.58% on the AffectNet-7, AffectNet-8 and RAF-DB datasets respectively.