Facial Expression Recognition Method Based on SpResNet-ViT
Weijun Gao, Lei Li, Huayang Zhao · 2022 2nd Asia-Pacific Conference on Communications Technology and Computer Science (ACCTCS) · 2022
Expression recognition in wild is susceptible to factors such as object occlusion and large pose. Existing facial expression recognition models lack the use of relationships between non-occluded regions. To address this problem, the SpResNet-Vit(ResNet with SoftPool-Vision Transformer) model consisting of a feature extraction module and a vision Transformer module is proposed. First, the original image is cropped and the feature extraction of the cropped image is performed using the optimized ResNet network; then, the attention to the non-occluded regions is enhanced using the self-attention module; finally, the internal relationships of the non-occluded facial regions are modeled using the global self-attention mechanism of the Vision Transformer. Our SpResNet-ViT is tested on three publicly available datasets, CK+, FERPlus and RAF-DB, and 98.6%, 88.1% and 87.3% accuracy rates were realized, which is significantly superior compared with current mainstream algorithms.