Aiming at the problems of insufficient expression of facial expression features, low recognition accuracy and many pa‐ rameters, an octave convolutional expression recognition method combined with attention feature fusion is proposed. The main inno‐ vation point is to introduce the attention feature fusion mechanism into the model to optimize the fusion of different scale features. Deep separable network is used to replace traditional convolution, which greatly reduces parameters. BN and PReLU are introduced to improve the stability and performance of the model. Experiments show that the accuracy of the model on CK+ and Fer2013 data sets is 98.91% and 74.03%, respectively, showing excellent generalization ability and accuracy.