부산대학교 도서관

Aiming at the problem of insufficient feature extraction and weak generalization ability in facial expression recognition, this paper proposes an improved hybrid attention mechanism model. Two convolution kernels of different sizes are first used in the channel attention module to extract the dual scale features of the image; Secondly, Softmax activation function is used to redistribute the channel attention weight and construct the relationship between different scale feature maps; Finally, replace the 3x3 convolution in the bottleneck layer with the module and insert the SAM module to improve the model's ability. The experimental results show that the accuracy of the proposed model in FER2013 and SFEW datasets can reach 73.89% and 56.42%, and possesses excellent generalization skills.