The Research On Face Expression Recognition Based On Self-Attention Mechanism

With the rapid development of computer science,especially artificial intelligence and computer vision,the use of machine perception to complete facial expression recognition(FER),thereby helping people better apply to real life scenes,has attracted more and more researchers’attention.In recent times,facial-expression recognition(FER)primarily focuses on images in the wild that include factors,such as face occlusion and image blur,rather than on laboratory images;this presents new FER challenges.This thesis studies how to better apply the self attention mechanism to face expression recognition in real scenes.The main work is as follows:(1)In order to recognize facial expression images in real environments including com-plex background,facial occlusion and other factors,a facial expression recognition method based on region enhanced attention network is proposed.Firstly,an attention-based region enhancement network is proposed to reduce the influence of external factors and enhance the robustness of expression recognition in real environments.Then,a channel-spatial attention fusion network is proposed to extract global features.Finally,the recognition degree of fa-cial expression images is improved by the combination of partition loss and cross entropy loss,thereby improving the recognition accuracy.The experimental results on public data sets RAF-DB,FERPlus and AffectNet show that the regional enhanced attention network has obvious advantages over other methods.(2)In view of the large parameters of the self attention mechanism,which can not be well applied to facial expression recognition tasks,a cross type dual attention method is pro-posed.The method comprises three parts—a cross-fusion grouped dual-attention mechanism to refine local features and obtain global information;a C~2activation function construction method is proposed,and the new C~2activation function is a piecewise cubic polynomial with three degrees of freedom,it not only requires less computation,but also has better flexibility and recognition ability,which can better solve the problems of slow running speed and neu-ron inactivation;and a closed-loop operation between the self-attention distillation process and residual connections to suppress redundant information and improve the generalization ability of the model.The recognition accuracy of the RAF-DB,FERPlus and AffectNet datasets were 92.78,92.02,and 63.58%,respectively.Experimental results showed that this model could provide a more effective solution for FER tasks.