In order to improve the accuracy of expression recognition and reduce the computational load of network, a deep learning-based expression recognition algorithm was proposed. The algorithm mainly preprocesses the image through the spatial transformation network (STN) in the early stage, and completes the image transformation processing required by the network. Then the super resolution (SR) algorithm is used to improve the overall image quality, so as to improve the overall feature information of the input image. Experimental results show that the proposed algorithm has a good performance in the accuracy of facial expression recognition, and can meet the requirements of accuracy and robustness of facial expression recognition algorithm.