Abstract: Aiming at the problems of a large number of parameters and the long training time of convolutional neural networks, we proposed
 a facial expression recognition method based on a lightweight attention residual network. Firstly, we rebuilt the model by using  the residual network as a skeleton, and  improved the model performance by reducing the number of layers and improving the residual module. Secondly, the depthwise separable convolution was introduced to reduce the number of model parameters and computational effort. Finally, the squeeze and excitation module of ReLU function was replaced by Mish function to adaptively 
adjust the channel weight. The model was validated by using the classical ten-fold cross-validation mode on two public datasets CK+ and JAFFE,  and obtained  accuracies of 98.16% and 96.67%, respectively. The experimental results show that the proposed method provides a better trade-off between model identification accuracy and complexity.

Key words: facial expression recognition, lightweight, residual network, depthwise separable convolution, attention mechanism