Research On Emotion Recognition Method Of Interactive Robot Based On Deep Learning

Human Computer Interaction(HCI)is a hot research direction in the field of smart home,and interactive robots use vision to recognize emotions is the most humanized detection method.Visual emotion recognition is the most humane detection method.In view of the poor performance of the current visual emotion recognition algorithm on face pose,illumination change and occlusion,and the need to integrate multi-modal information recognition in the future,this thesis conducts research on emotion recognition based on Transformer neural network.The main research content of this thesis is as follows:First,a network model for facial expression and emotion recognition is constructed based on the visual Transformer.On the basis of the current algorithm research of the visual Transformer model,the structure of the general framework of the visual Transformer is improved for emotion recognition tasks.The post-normalization structure is used to construct the components of the model,and then different self-attention mechanism modules are mixed to extract features,and a hybrid visual Transformer model CS Former is proposed.Conduct experiments on datasets CK+,FERPlus,RAF-DB and self-made datasets FER-T to analyze key information such as model accuracy.Compared with the CS Former model and the mainstream model,the CS Former model can reduce the amount of parameters by 50% while sacrificing only 2% of the accuracy rate,speed up the calculation of the model,and reduce the memory occupied by the model.Secondly,research on the lightweight method based on the visual Transformer emotion recognition model.On the basis of the emotion recognition model,the component calculation amount of the CS Former model is reduced by merging Token.Aiming at the problem of high computational complexity of the self-attention mechanism module in the CS Former model,this thesis proposes a method of repeatedly using selfattention weights to reduce the amount of calculation of the model.Then design the knowledge distillation method to train the lightweight model,and conduct experiments on the data sets CK+,FERPlus,RAF-DB and self-made datasets FER-T.The results show that the optimized CS Former model can reduce 36% of floating-point calculations and speed up the calculation of the model while only sacrificing 0.4% of the accuracy rate.Finally,an interactive prototype of emotion recognition is designed based on the CS Former model.The interactive prototype was developed using the Unity engine.First use Blender to make interactive anthropomorphic expression animations,then use Unity to design the state machine of expression animations,and then embed the emotion recognition detection part into the interactive prototype.After the interactive prototype uses the camera to collect data,detect the human face and expression emotion,and then change the anthropomorphic expression animation state machine,so that the interactive prototype can effectively interact with the user.Experiments verify that the interactive prototype can interact effectively under the conditions of face posture changes,illumination changes and occlusions.