Occlusion-aware Multi-scale Attention Consistency Network for Facial Expression Recognition

Facial expression recognition (FER) faces difficult challenges in real-world scenarios, especially with occlusion and pose changes. Previous approaches typically utilize face key points to improve FER performance. However, face key point detection can be inaccurate, especially in the presence of occlusions and pose variations, and manually labeling face key points is a complex and time-consuming task. To address this issue, we introduce a simple and efficient Occlusion-aware Multi-scale Attention Consistency Network (OMAC-Net), which can capture global and local fine-grained facial expression representations. The proposed OMAC-Net consists of a Low-level Feature Random Destruction (LFRD) module and a Multi-scale Attention Consistency Constraint (MACC). Specifically, LFRD is used to obtain rich local fine-grained information and retain important patch edge information and correlation information between patches. Then MACC further constrains the model's attention, enabling the model to focus on discriminative features. The experimental results on three in-the-wild FER convincingly demonstrate the superiority of OMAC-Net compared to state-of-the-art methods. Our code and models will be available at this https https://github.com/lfw147/OMC.