Research On Facial Expression Recognition Based On Convolutional Neural Network

In recent years,with the development of the information age,facial expression recognition has become an important topic in the field of computer vision,which is widely used in fields such as human-computer interaction,medical assistance,online education,and safe driving.In view of the current challenging problems in facial expression recognition,this paper combines deep neural networks to conduct research from feature extraction,database limitations,convolutional neural network optimization,loss function improvement,etc.The main work content is as follows:(1)Considering that a single feature is not enough to represent face information,this paper proposes an expression recognition algorithm based on a dual-channel convolutional neural network based on local binary pattern(LBP)and gradient features.Firstly,the image is preprocessed and the corresponding LBP image and gradient image are generated.Then input it into two separately improved Inception-Res Net-V2 networks.Channel 1 inputs LBP images to extract facial texture information and capture subtle facial movements;channel 2 inputs gradient images to extract facial structure and edge features.The two features complement each other to more efficiently and comprehensively characterize facial features.Finally,they are weighted and fused together,and Softmax is used for classification.We conducted a lot of experimental analysis and verification on CK+,JAFFE,FER2013,Oulu-CASIA and self-collected NCUFE data sets,and the results proved that our model is more competitive.(2)It is invariant to posture and illumination for depth information.This article uses Kinect V2 sensor to collect a brand new RGB-D facial expression image database.At the same time,it is aimed at the efficiency problems caused by the complex structure of most current convolutional neural networks and the excessive number of layers.We propose a lightweight convolutional neural network(Efficient Net-B0)based on the attention mechanism for expression recognition in RGB-D databases.The network uses the method of compound model expansion to uniformly scale the depth,width,and resolution,which reduces the weight of the model and improves the overall performance.In addition,we have combined Softmax and Island functions to redesign the loss function of the network to further improve the feature discrimination effect.We have conducted a lot of experiments and analysis on the self-collected RGB-D expression database,and the results have proved the effectiveness of our network.