Facial emotion recognition based on AlexNet with large margin cosine loss function

In order to improve the accuracy of facial emotion recognition, a Large Margin Cosine Loss(LMCL) function is proposed combined with convolutional neural network (CNN) for emotion recognition. The LMCL function is a loss function which could increase the differences of features among samples from different classes(inter-class), meanwhile reduce the differences of features among samples from same classes (inner-class). A reliable and simple neural network Alexnet is customized and trained for this task with dataset FER-2013, a wildly used dataset for emotion recognition. Compared with the neural network with original loss function cross entropy with softmax(The following is referred to as the following softmax loss function), the accuracy of neural network model is improved by 21.57% with LMCL function in experiment. As well as the error rate of different classes recognition is reduced.