TY  - JOUR
AU  - WANG, Kunxia
AU  - YU, Wancheng
AU  - HU, Yuxia
PY  - 2024
TI  - Facial expression recognition in Swin Transformer by embedding hybrid attention mechanism
JO  - Journal of Northwest University (Natural Science Edition)
SN  - 1000-274X
SP  - 168
EP  - 176
VL  - 54
IS  - 2
AB  - Facial expression recognition is an important research domain in psychology that can be applied to many fields such as transportation, medical care, security, and criminal investigation. Given the limitations of convolutional neural networks (CNN) in extracting global features of facial expressions, this paper proposes a Swin Transformer method embedded with a hybrid attention mechanism for facial expression recognition. Using the Swin Transformer as the backbone network, a hybrid attention module is embedded in the fusion layer (Patch Merging) in the model of Stage3, which can effectively extract global and local features from facial expressions. Firstly, the hierarchical Swin Transformer model can effectively obtain deep global features. Secondly, the embedded hybrid attention module combines channel and spatial attention mechanisms to extract features in the channel dimension and spatial dimension, which can attain better local features. At the same time, this article uses the transfer learning method to initialize the model network weights, thereby improving the recognition performance and generalization ability. The proposed method achieved recognition accuracies of 73.63%, 87.01%, and 98.28% on three public datasets (FER2013, RAF-DB, and JAFFE) respectively, achieving good recognition results.
UR  - https://doi.org/10.16152/j.cnki.xdxbzr.2024-02-003
DO  - 10.16152/j.cnki.xdxbzr.2024-02-003