GFFT: Global-local feature fusion transformers for facial expression recognition in the wild

Facial expression recognition in the wild has become more challenging owing to various unconstrained conditions, such as facial occlusion and pose variation. Previous methods usually recognize expressions by holistic or relatively coarse local methods, but only capture limited features and are susceptible to be influenced. In this paper, we propose the Global-local Feature Fusion Transformers (GFFT) that is centered on cross-patch communication between features by self-attentive fusion. This method solves the problems of facial occlusion and pose variation effectively. Firstly, the Global Contextual Information Perception (GCIP) is designed to fuse global and local features, learning the relationship between them. Subsequently, the Facial Salient Feature Perception (FSFP) module is proposed to guide the fusion features to understand the key regions of facial features using facial landmark features to further capture face-related salient features. In addition, the Multi-scale Feature Fusion (MFF) is constructed to combine different stages of fusion features to reduce the sensitivity of the deep network to facial occlusion. Extensive experiments show that our GFFT outperforms existing state-of-the-art methods with 92.05% on RAF-DB, 67.46% on AffectNet-7, 63.62% on AffectNet-8, and 91.04% on FERPlus, demonstrating its effectiveness and robustness.

Automated summary

In this paper, the Global-local Feature Fusion Transformers (GFFT) method is proposed to address the challenges of facial expression recognition in the wild, such as facial occlusion and pose variation. The GFFT approach utilizes self-attentive fusion to enable cross-patch communication between features, effectively capturing global and local information. Experimental results demonstrate the effectiveness and robustness of GFFT, outperforming existing state-of-the-art methods on multiple datasets.