In view of the fact that the proportion of face in the whole image is small and the face posture is skew, which leads to the low accuracy of facial expression recognition, this paper proposes a two-branch neural network based on attention mechanism and length feature of facial landmark. The first branch is based on the modified VGG19 and the integration of the convolutional block attention module (CBAM), which focuses on the extraction of important features of expression. The second branch extracts 68 feature points from the face image, and calculates the Euler distance between these feature points to acquire the length information of 2278 dimensions, and is finally normalized and fed into the full connection layer. The network composed of the first branch and the second branch is trained by joint fine-tuning, where L2 regularization is added to prevent overfitting. The proposed method is tested and verified on Jaffe, CK + and FER +, and the highest accuracy is 93.48%,96.12% and 84.73% respectively. The performance of the proposed method is more advance than many existing algorithms, which shows that the proposed model has good accuracy and robustness.
Paper
Full text
Expression Recognition Based on Attention Mechanism and Length Feature of Facial Landmark
OpenAlex · Face and Expression Recognition · 2021
Abstract
In view of the fact that the proportion of face in the whole image is small and the face posture is skew, which leads to the low accuracy of facial expression recognition, this paper proposes a two-branch neural network based on attention mechanism and length feature of facial landmark. The first branch is based on the modified VGG19 and the integration of the convolutional block attention module (CBAM), which focuses on the extraction of important features of expression. The second branch extracts 68 feature points from the face image, and calculates the Euler distance between these feature points to acquire the length information of 2278 dimensions, and is finally normalized and fed into the full connection layer. The network composed of the first branch and the second branch is trained by joint fine-tuning, where L2 regularization is added to prevent overfitting. The proposed method is tested and verified on Jaffe, CK + and FER +, and the highest accuracy is 93.48%,96.12% and 84.73% respectively. The performance of the proposed method is more advance than many existing algorithms, which shows that the proposed model has good accuracy and robustness.