Research On Facial Expression Recognition Method In Natural Scenes Based On Deep Learning

Facial expression is one of the important symbols of body language.The subtle changes of expressive muscles produce a rich variety of expression forms,which can convey the individual’s inner state of mind and emotional state.With the development of computer vision technology,facial expression recognition has become an important research direction in the field of vision and has applications in affective computing,human-computer interaction,and security surveillance,among others,and has significant social and commercial value.Although traditional laboratory settings often achieve good results for facial expression recognition technology,it has many limitations in complex and changing natural environments,mainly in difficulties in feature extraction,strong data dependence,and poor expression recognition performance.To overcome these shortcomings,deep learning-based methods have emerged in recent years,which abandon the traditional manual design of feature extractors and realize endto-end training and powerful feature expression capabilities.Although deep learning technology performs well in facial expression recognition applications in natural scenes,it is still limited by factors such as non-frontal face poses,partial occlusion,and complex backgrounds,and further improvements in recognition accuracy are needed.Therefore,this paper proposes two improved methods based on convolutional neural network models and Transformer models to improve the recognition rate of facial expressions in natural scenes.The main research work of this paper is as follows:(1)In view of the fact that facial expression recognition in natural scenes is easily affected by occlusion,non-frontal posture and other factors,and the representation bottleneck problem of traditional convolutional neural network,this paper designs a facial expression recognition method based on improved rank expansion network(ReXNet-TS),using a combination of 3D interactive attention and semantic aggregation,effectively improves the accuracy of facial expression recognition in natural scenes.Specifically,the following work has been completed on the basis of the rank expansion network:the expression feature analysis ability of the rank expansion network is improved by cutting and optimizing the fully connected layer and multilayer scale feature learning;a three-dimensional interactive attention mechanism is designed to solve the pose change,For issues such as lighting effects and complex backgrounds,this mechanism adaptively assigns different weights to its key features,strengthens recognizable key features,and suppresses redundant features,so that the network model can capture more discriminative features;at the same time,In order to strengthen the information interaction between features,a channel confusion module is designed to make full use of channel features at the same stage;finally,a semantic interaction module is designed to use multi-layer features more efficiently and reduce the loss of semantics in shallow and middle layers.Influence.Experiments show that the ReXNet-TS method has achieved recognition rates of 88.89%,89.53%and 62.22%on the facial expression datasets RAF-DB,FERPlus and AffectNet-8 respectively,confirming the superiority of the method.(2)Aiming at the insufficient global feature modeling ability of traditional network models in natural scenes and the uneven quality of expression samples,this paper designs a facial expression recognition method(CSWin-SA)based on an improved visual Transformer.The CSwin Transformer is different from the convolutional neural network.It uses a cross-shaped window to control the feature receptive field,so it can expand the field of view without increasing the depth of the network.On this basis,the discriminative facial expression is further enhanced by using 3D interactive attention to optimize the underlying components of the network.Secondly,in order to obtain a more ideal expression sample input,a spatial transformation module is used to transform the input expression samples.At the same time,during the dynamic training process,Adaface face loss is weighted and fused with cross-entropy loss to alleviate the impact of low-quality expression samples on model performance.Experiments show that the CSWin-SA method achieves recognition rates of 89.50%,90.23%and 62.58%on the facial expression datasets RAF-DB,FERPlus and AffectNet-8,respectively,confirming the advanced nature of the method.Finally,combined with PyQt5,a facial expression recognition application system is designed,and the facial expression recognition performance test of the system in natural scenes is carried out.The test results show that the recognition performance in natural scenes is better.