Research On Facial Expression Recognition Algorithm Based On Transformer

Facial expression contains rich emotional information,which can directly express the current emotional state of human beings.With the progress of science and technology,people’s demand for intelligence is gradually increasing.As a key part of intelligence,FER task has attracted much attention.At present,most algorithms for facial expression recognition are based on convolutional neural networks.However,the convolutional filter in CNN relies heavily on spatial locality and cannot learn the global features of facial expression at the beginning of the model.The model needs to stack the convolutional layers to make the network perceptive field larger,and finally learn the global features.In most cases,this process will lead to the loss of input information,increase the amount of computation,and also exist the problem of vanishing gradient.Due to the problems of CNN’s own characteristics,many researchers begin to focus on how to apply Transformer model to image tasks to try to make up for the shortcomings of CNN.This paper conducted the following research:(1)This paper proposes an Patch-Range Attention facial expression recognition algorithm(PVS)based on Visual Transformer.PVS integrates three different attention mechanisms.Firstly,PRA module is used to realize the attention mechanism within the scope of image patches,which can extract the local features of image patches.Then,the global attention mechanism used by vision Transformer can be used to deeply extract image features.Finally,the SE module can realize the channel attention mechanism and complete the final feature extraction.Through these three different attention mechanisms,our model is very robust for facial expression recognition.We conduct experiment on PVS on different facial expression recognition data sets,and the results show that the PVS model has strong performance.(2)We propose a facial expression recognition algorithm(GTVi T)with vision Transformer integrating global tokens.Visual Transformer simply divides the facial image into fixed size image patches,which will inevitably lead to information loss in adjacent areas between adjacent image patches.Meanwhile,Visual Transformer directly convert image patches into sequences,which will lose a lot of internal information of image patches.Therefore,we propose a transformer model integrating global tokens.fusion module.By integrating global tokens and image patch tokens,the semantic features of facial expression can be enhanced.In this way,the visual Transformer can enhance its attention to key facial expression features and learn the relationship between each key areas and image patches.We conduct experiments on GTVi T on three facial expression recognition data sets.The results show that GTVi T can adaptive learn the key regions for expression classification,and has a strong ability for expression classification.(3)We design and develop a set of real-time face detection,face recognition and facial expression recognition system,in order to explore the performance of PVS in the actual products.A total of three modules have been developed to complete the system requirements,which are face detection module,facial expression recognition module and face recognition module.We do some research on face detection and face recognition,and apply the model in this paper to our system.The results show that our model can achieve good recognition effect in the actual production environment.