Research and Analysis of Learning Rate on Face Expression Recognition

As an important hyperparameter in facial expression recognition, learning rate determines whether the facial expression recognition function can converge to the local minimum and when it converges to the minimum value. With the increasing complexity of the facial expression recognition network model, it is particularly difficult to find a suitable learning rate, because the size of different parameters varies greatly and needs to be adjusted throughout the training process, which not only consumes a lot of time. Sometimes, it is not possible to adjust to the optimal, which makes it difficult to improve the accuracy of facial expression recognition. Therefore, in this paper, through the research and analysis of the facial expression recognition rate, we propose face expression recognition based on the Adabound algorithm and compare it with SGD (Stochastic Gradient Descent) and Adam on the face expression recognition network model use CK+ (Cohn-Kanade) dataset. The Adabound adaptive optimization algorithm achieved high accuracy of 96.97% and 100% on ResNet18 and ResNet34 respectively. Compared with recent articles, our model based on the Adabound algorithm has reached a very high level in the CK+ dataset in terms of the effect of a single model.