Aiming at the problem of redundant features when using deep learning to extract facial expression im? age features，an improved Xception facial expression recognition network based on multi-layer perceptron（MLP）is proposed. In this model，the features extracted from the Xception network are input into the multi-layer perceptron for weighting，the main features are extracted，and the redundant features are filtered out so that the recognition accu? racy is improved. First，the image is scaled to 48*48，then the data set is enhanced，and these processed images are fed into the network model proposed in this paper. A comparison of ablation experiments show that：The correct rec? ognition rates of this model on the CK + dataset，JAFFE dataset，and MMI dataset are 98.991%，99.02% and 80.339% respectively. The correct recognition rates of Xception model on the CK + dataset，JAFFE dataset and MMI dataset are 97.4829%，90.476%，and 74.0678%，respectively. The correct recognition rates of the Xception + 2lay model on the CK + dataset，JAFFE dataset and MMI dataset are 98.04% and 74.0678%，84.06%，and 75.593%，re? spectively. By comparing the above ablation experiments，the recognition accuracy of this method is significantly bet? ter than the Xception model and the Xception + 2lay model. Compared with other models，the effectiveness of this model is also verified.