Enhancing Real-World Facial Expression Recognition: A Deep Learning Approach based on Attention Mechanisms

Occlusion and pose variation are the biggest interference factors in real-world face expression recognition (FER). Faces are likely to be occluded by items such as sunglasses, ducktails, and scarves, which can significantly change the visual appearance of the face, while varied poses can lead to partial information loss and inaccurate localization. In this paper, we propose a new deep learning network model for facial expression recognition, firstly, an attention mechanism module is added on top of the residual neural network, which includes spatial and channel attention modules for suppressing unimportant feature information and highlight the key information. Second, a regional attention mechanism is added for adaptively extracting feature vectors from localized unobstructed facial regions, which are combined with global feature information to learn more comprehensive and detailed potential key information. The proposed method achieved good results on two facial expression datasets, FER2013 and RAF-DB, obtaining an accuracy of 74.37 and 89.12, respectively.