Abstract: Aiming at the problems of low accuracy, complex model and large amount of calculation in current facial expression recognition, an improved facial expression recognition method based on octave convolution was proposed in this paper. The improved octave convolution is used for feature extraction, which improves the extraction effect of detailed feature, decline the redundancy in the feature map, and reduce the amount of calculation without increa-sing the parameters, so as to improve the feature extraction performance; DyReLU activation function was used to enhance the learning and expression capabilities of the model. The parameters are reduced using an adaptive mean pooling downsampling layer instead of a fully connected layer; the model is pre-trained on a large dataset, and then the model performance verification experiment is performed on the FER2013, FERPlus and RAF-DB datasets. The experimental results show that the trained model weights is only 10.4 MB, and the accuracy of the model on the expression recognition datasets FER2013, FERPlus and RAF-DB are 73.53%, 89.58% and 88.50% respectively. Compared with many current models, the OCNN model had higher accuracy and lower computing resource consumption, which fully proved the effectiveness of this model.