Abstract: Aiming at a low accuracy of facial expression recognition and the susceptibility to noise and other factors under natural conditions, this paper proposes a facial expression recognition method that incorporates an improved deep and wide residual network with attention mechanism. A wide residual module structure is formed by increasing the channel number ofresidual units, which effectively alleviates the problem of gradient disappearance caused by excessive network layers.In order to understand facial expressionfornetworks, a compressed and adaptive correction network module is introduced.In order to alleviateover-fitting of the model, the order of the residual units is improved.The original imagesare processed through the improved random erasure method to further strengthen the generalization ability of the model. Experimental results show that the accuracy of the model on fer2013, ck+ data set and JAFFE data set are 72.49%, 99.29% and 94.87%respectively.Compared with other methods, the model proposed in this article has a much higher recognition accuracy and, at the same time,a better robustness.