Research On Facial Expression Recognition Based On Key Regions

As an important branch of computer vision,facial expression recognition has broad application prospects in intelligent monitoring,human-computer interaction and other fields.In recent years,more and more researchers use convolutional neural network to study expression recognition,and have made great achievements.At present,most of the models used for expression recognition have a large number of parameters,which cannot meet the requirements of mobile devices and embedded devices.Therefore,this thesis studies the lightweight model mini＿Xception,improves the structure of the lightweight model,extracts the features of key regions,makes the model pay more attention to the important regions related to expression,and realizes accurate expression recognition.Aiming at the problem that convolutional neural network cannot focus on the most significant region of expression recognition,so it cannot effectively extract the key feature information.This thesis studies the structure of the basic model mini＿Xception,and extracts the feature of the whole face by introducing the SE-Net attention module.The attention module can adaptively learn the importance of different feature channels,increase the attention of the model to the key regions related to expression in the whole face,and obtain the whole facial expression recognition network based on attention mechanism(WFA-CNN).Experimental results show that the method can effectively improve the accuracy of facial expression recognitionTo solve the problem that WFA-CNN still ignores the subtle feature changes of some local facial expressions,this thesis constructs the facial expression recognition networks of local key regions,LER-CNN,LMR-CNN and LNR-CNN based on FACS,to realize the learning of the detailed features of the constructed regions.By learning the local details of facial expressions,each key region model has the ability to recognize the local details of facial expressions.In the process of constructing the local key region model,this thesis also proposes a method to locate the key region by adjusting the parameter α,which can reduce the missed and false detection rate of the key region and improve the detection ability of the local key region.Finally,in order to complement the local key region expression recognition networks and the whole facial expression recognition network based on attention mechanism,and improve the accuracy of expression recognition.In this thesis,the WFA-CNN is fused with local key region expression recognition networks by using the method of model fusion.By using different fusion strategies,a multi-models fusion expression recognition network is constructed.The experimental results show that the model fusion method can capture and utilize the important regions related to expression in the face,and display these important regions of each model through CAM visualization.Experiments on CK+ dataset and JAFFE dataset,show that the recognition accuracy of the model fusion method is effectively improved.