Facial Expression Recognition Based On Residual Network

Facial expressions can properly express inner emotions. The differences between expressions also make feature extraction the most important part of expression recognition. Among all the deep learning network models, the residual network put forward by Kaiming He et al. dose better in network training. Therefore, on the base of the residual network, this paper will replace the convolution in the residual block by Pyramid Convolution. At the same time, the attention module is introduced to redistribute the weight parameters of channel and spatial dimensions, also the normalization operation is improved. The results show that the accuracy on FER2013 and CK + data sets reached 72.276% and 96.970% respectively. In comparison to the unmodified model, the error rate is reduced by 2.867% and 5.758%. The improvement of the model is proved to be effective.