Abstract: A video expression recognition method based on facial motion units and temporal attention is proposed to address the problem of inconsistent expression intensity in video sequences, which is difficult to extract features effectively by a long short-term memory network(LSTM). Firstly, we introduce a temporal attention module based on convolutional LSTM(ConvLSTM) to model the video sequences temporally,which can reduce the dimensionality while retaining the rich feature information of face images. Secondly,we propose a face image segmentation rule based on facial motion units to solve the problem that it is difficult to define the active regions of facial expressions. Finally, we embed a label correction module in the model to solve the problem of sample uncertainty in the data set under natural conditions. The experimental results on MMI, Oulu-CASIA and AFEW datasets show that the number of model parameters of this method is lower than that of the published mainstream models, and the average recognition accuracy on the MMI dataset is 87.22%, which is higher than that of the current mainstream methods, and the overall effect is better than that of the current representative methods.