基于分离混合注意力机制的人脸表情识别

近年来，人脸表情识别研究因机器学习的引入取得了显著的进步，但由于光照变化、人脸姿态等原因使得人脸表情识别准确率一直不高。提出了一种基于混合注意力机制网络的方法，把通道注意力和空间注意力机制分离，增强网络的通道特征和跨通道相关性学习能力。在注意力机制后增加软阈值机制以抑制噪声，在网络损失函数中迭加类中心损失减小类内差异，通过预处理提高网络的泛化能力。实验结果表明，该方法在人脸数据集CK+和fer2013上准确率比主流方法更高，所用参数更少，收敛性更好。

Facial expression recognition based on separate hybrid attention mechanism

In recent years,machine learning has made significant progress in facial expression recognition.However,the accuracy of facial expression recognition is not high due to illumination variation,face pose and other reasons.A method based on hybrid attention mechanism network is proposed,which separates channel attention from spatial attention to enhance learning ability of the network on channel characteristics and cross channel correlation.A soft thresholding mechanism is added after the attention mechanisms to suppress noise,center loss is added to the network loss function to reduce differences within classes,generalization ability of the network is improved by preprocess.Experiment results show that compared with those of mainstream methods,the accuracy of the proposed method is higher on face dataset CK+ and fer2013,the parameters are fewer,and the convergence is better.