Research On Video Facial Expression Recognition Based On Deep Learning

As one of the research hotspots,facial expression recognition has been widely used in the fields of autonomous driving,social networking,and classroom teaching.Facial expression recognition involves the cross-integration of multiple disciplines,such as computer science,biology,psychology,etc.,and is a valuable research topic.Due to the subtlety of expressions,the expressions on faces are not easy to distinguish,so real-time facial expression recognition is still a big problem.At the same time,this paper mainly focuses on video data,which is more complicated than single-frame images.Therefore,based on the above problems,this paper optimizes the design based on the traditional VGG-16+LSTM network framework,aiming to improve the accuracy of video facial expression recognition.The specific work is as follows:(1)Aiming at the problem of inaccurate expression feature extraction by traditional networks,this paper proposes Feature Enhancement Convolutional Networks(FECNN),which is studied from the perspectives of single-frame feature enhancement and inter-frame feature enhancement to improve the accuracy of video expression recognition.rate purpose.First,a 7×7 convolutional layer operation is extended in the middle layer of VGG-16 to extract shallow facial expression features,and fused with deep features to increase the spatial information of facial expressions;then,the expansion rate is applied in the last layer of VGG-16 It is an atrous convolution of 2,which reduces the information loss while increasing the receptive field of the convolution operation.Next,the Squeeze-Excitation mechanism is used to give weights to the facial expression feature channels to improve the accuracy of facial expression single-frame features.Finally,a Self-attention mechanism is introduced to give weights to video frames according to the correlation between video frames to improve the accuracy of multi-frame features of facial expressions.The theoretical idea has been verified and compared on the AFEW dataset,CK+ dataset,SFEW dataset,and FER2013 dataset,which confirms the superiority of the model.(2)Aiming at the problem that the inter-frame attention mechanism in the feature-enhanced convolutional network is not suitable for processing long sequences and the information extraction is not rich,this paper adopts the multi-head attention mechanism in the transformer to replace the interframe attention mechanism,and aims at the information caused by the deviation of attention weights.The dropout problem proposes the Multi-Head Prior Attention Mechanism(MHPAM).The model is simulated and verified on the AFEW dataset and the CK+ dataset,which confirms that MHPAM can improve the accuracy of facial expression recognition.(3)Aiming at the problem that the inter-frame attention mechanism in the feature-enhanced convolutional network is not suitable for processing long sequences and the information extraction is not rich,a multi-head attention mechanism in the transformer is proposed to replace the inter-frame attention mechanism.A new Multi-Head Prior Attention Mechanism(MHPAM)is introduced,which combines the output feature map of the multi-head attention mechanism with the VGG-16 output feature map,and uses the VGG-16 output feature map for feature supplementation and attention.Weight guidance greatly enriches the features.The model is verified by simulation on AFEW dataset and CK+ dataset,which confirms the superiority of the model.