Differential Residual Learning for Facial Expression Recognition

Abstract

The facial expression recognition algorithm based on convolution neural network (Convolutional Neural Network) still has the problem of imperfect feature point extraction. Therefore, a facial expression recognition algorithm based on the combination of the detail feature extraction network and pre-training model is proposed. According to the emotional face and neutral face of the residual of the images, learn the residual to extract features, and make the feature information accurate. Then, the face feature markers generated by the feature extraction network are loaded into the pre-training model respectively for classification and recognition. By grouping and crossing experiments on datasets CK+ and FER2013, the average accuracy of face recognition is 95.74% and 73.11%, respectively. Compared with the state-of-the-art recognition model, this method is effective in facial expression recognition to some extent.