Research On Expression Recognition Algorithm Based On Deep Learning

As a means of emotional expression,facial expressions including rich emotional information and psychological activities,play an important role in human language communication.The main purpose of facial expression recognition is to enable the calculation to perform feature extraction analysis on human expressions autonomously,to classify and understand according to human cognition and thinking,to think and infer through prior knowledge of human emotional information,to identify the emotions of others,such as happiness and anger,ultimately achieve human-computer interaction.In recent years,deep learning algorithms have performed well in many fields.Based on this,this paper uses deep learning to do the following research on facial expression recognition.Aiming at the characteristics of facial expression data,a network structure FER-Net for expression recognition was designed.The network is divided into two parts:feature fusion network and classification network.In the feature fusion network,the feature maps with different scales at different CNN levels are simultaneously used,and the bilinear interpolation method is used to upsample the small scale feature map,and then merge with the large scale feature map to obtain strong semantics.The feature map of the information enhances the deep features of the network;in the classification network part,like most networks,the classification result is output through a softmax function through convolution and pooling operations.The accuracy of 73.06%and 85.40%was obtained by experiments in FER2013 and RAF-DB,respectively,which verified the feasibility of FER-Net.In order to further improve the recognition effect of facial expressions,a attention mechanism is introduced in FER-Net to explicitly model the interdependence between feature channels,that is,the network can use the global information to selectively enhance the beneficial feature channels and suppress useless feature channel;at the same time,after the last convolutional layer of FER-Net,the second-order global covariance pooling is used to replace the first-order global average pooling,which captures richer deep feature information in the network and improves network expression ability and generalization ability.Through the improvement of FER-Net,the recognition accuracy of 73.70%and 85.85%was obtained in FER2013 and RAF-DB respectively,which significantly improved the classification performance of the model.