Facial Expression Recognitionical Based On Convolutional Neural Networks

Facial expressions are the most direct non-verbal signals of human emotions,and by identifying such signals,a person’s genuine emotions can be roughly inferred.Therefore,facial expression recognition technology has great application value in human-computer interaction,criminal interrogation,traffic safety,and medical monitoring.In recent years,with the rapid development of deep learning,people can use convolutional neural networks’ excellent feature learning ability for facial expression recognition.Although the convolutional neural network has achieved great success in facial expression recognition,the convolutional neural network needs to extract sufficient expression features,which quickly leads to the relatively average accuracy of the model’s face expression recognition.In addition,in practical application scenarios,due to nonlinear factors such as gender,age,and ethnic background,there are problems of significant intra-class changes and slight differences between facial expressions.In view of these problems,this thesis conducts in-depth research on the facial expression recognition algorithm based on the convolutional neural network,and the main research work includes the following points:1.This thesis proposes a facial expression recognition algorithm based on the multi-scale adaptive parallel integrated network.The algorithm extracts multi-scale expression features through three different subnetworks to solve the problem of insufficient feature extraction by convolutional neural networks.At the same time,the algorithm embeds a channel-spatial attention module and a dilated convolution module in the sub-network LPF-Net and LSF-Net to enhance the feature extraction ability of the sub-network.In addition,the algorithm replaces the ordinary convolutions in the sub-networks LPF-Net and LSF-Net with depthwise separable convolutions,and selects the first three stages of the Swin-T network as the sub-network PST-Net to reduce the number of parameters of the parallel integration network.The proposed method conducted extensive experiments on FER2013 and RAF-DB datasets,achieving facial expression recognition accuracy of 74.94% and 88.17%,respectively.Experiments show that the expression recognition performance of the proposed method is better than most expression recognition methods.2.This thesis proposes a multi-branch network face expression recognition algorithm based on gender constraints.In the data preprocessing stage,the algorithm redivides the expression dataset into strong and weak similarity sets under gender constraint through gender pseudo-label labeling,VGG19-based expression feature extraction,and the K-means clustering method reducing the influence of gender factors on expression recognition.At the same time,the algorithm embeds the channel attention mechanism into the strongly similar branch network in the feature extraction stage.The mechanism can enhance the feature representation of proper channels according to the channel information,thereby effectively reducing the occurrence of misidentification of expressions in vital similarity centers.In this thesis,extensive experiments were performed on CK+,FER2013,and RAF-DB datasets.Experiments show that compared with some advanced methods,the network model proposed in this thesis has better recognition performance.