This paper proposes a Dual-branch Feature Fusion(DFF)method to enhance the robustness of head posture in facial expression recognition,addressing the issue of low recognition performance caused by head posture interference.Firstly,in the expression branch,high-dimensional rough semantic features are extracted using the ResNet18 backbone network.Then,the Spatial Feature Enhancement(SFE)module is employed to facilitate information interaction among high-level semantic features at the spatial level,thereby improving the expression feature extraction capability.Meanwhile,in the head pose branch,head pose features are extracted using the Head Pose Feature Extraction(HPFE),which is pre-trained on the head pose dataset 300W_LP with fixed weights.Finally,the expression features in the expression branch and the head pose features in the head pose branch are fused element-by-element to attain complementary information and establish a pose-robust emotional representation.The proposed method is evaluated on two widely-used datasets:RAF-DB dataset and FERPlus dataset.On the Pose Variation test set,the recognition accuracy of the two head poses(Pose＞30° and Pose＞45°)is 89.98％and 89.96％on the RAF-DB dataset,and 89.20％and 87.94％on FERPlus dataset,respectively.The experimental results show that the method proposed in this paper improves the accuracy of facial expression recognition in images under head posture interference,which is of great significance for research on facial expression recognition in natural environments.
万方数据
维普
