Facial Expression Recognition Based On Residual Neural Network

Facial expressions are an important part of human emotional signals and an important means of communicating emotions between people.With the development of artificial intelligence,facial expression recognition has become an important research topic in the field of computer vision.Methods based on deep learning have achieved certain results in facial expression recognition.However,with the development of convolutional neural networks and the increase of network depth,many of these methods still have problems such as network degradation,loss of feature information,single feature extraction,and low recognition accuracy.In order to solve the above problems,we have carried out the following aspects of research:(1)In this paper,the residual neural network is selected as the main core network,but when the gradient of the Re LU(Rectified Linear Unit)activation function used in the residual module is 0,it will inactivate neurons and lead to the loss of feature information.To solve this problem,we use the Mish activation function instead,and the slight tolerance for negative values in the Mish activation function improves the gradient flow.The designed residual neural network can not only extract deep features,but also retain shallow features,which can effectively prevent network degradation problems.(2)On the basis of the residual neural network,the Inception module is introduced,and the1 × 1 convolution kernel and multi-scale feature aggregation are introduced into the Inception module,which can not only obtain multi-scale feature information,but also reduce the dimensionality of convolution,effectively reduce the complexity of calculation and speed up the calculation.Moreover,the features extracted by the residual module and the features extracted by the Inception module are combined to form a dual-channel network,which makes the features extracted by the network model richer.(3)Instead of abandoning the traditional feature extraction algorithm,the features extracted by the traditional LBP algorithm are combined with deep learning features,which can not only enable the network to obtain richer feature information,but also alleviate the impact of lighting and picture rotation,and further improve the accuracy of expression recognition.Through verification experiments on the public datasets CK+(The extended Cohn-Kanade Dataset)and KDEF(Karolinska Directed Emotional Faces),the problems of network degradation,insufficient extracted feature information and low recognition accuracy were effectively solved,and the recognition accuracy of 96.429% and 95.855% on the two datasets was achieved,respectively.