Research On Emotion Recognition Based On Lifelong Learning

Human Computer Interaction(HCI) has become a hot research direction in the field of artificial intelligence,but traditional HCI can only achieve accurate and mechanical interaction functions,and cannot understand human emotions.Therefore,in order to realize HCI with natural and temperature,emotion recognition is essential.Current emotion recognitions are mostly based on a single task,i.e.training a separate model for each task.This model has the problems of low model training efficiency,unreliable recognition,and incomplete emotion recognition.In view of the above problems,on the basis of analyzing the existing emotion recognition methods,this thesis proposes a research method for emotion recognition based on lifelong learning.The use of lifelong learning can take into account the characteristics of multi-task recognition and positive transfer of knowledge,break the limitations of emotion recognition from a single source of information,and achieve more efficient,reliable and comprehensive emotion recognition.The main research works of this thesis are as follows:1.Create own dataset.Aiming at the defect that the existing emotion recognition datasets are independent and unrelated,this thesis uses the emotion induction method of video materials to design an emotional acquisition experimental scheme,and two kinds of data,EEG signal and face image,are collected at the same time.The two kinds of data belong to the same emotion and can be verified by each other.2.Based on the emotion recognition of own database,the effectiveness of the emotion recognition method based on lifelong learning is preliminarily verified.Based on the principle of progressive neural network in lifelong learning,the Bi-ProgNet network is constructed,which can simultaneously perform emotion recognition on facial expression data and EEG signal data,and the recognition accuracy is 91% and 87.5%,respectively.The experimental results demonstrate the effect of progressive neural network knowledge forward transfer and the feasibility of applying lifelong learning to emotion recognition.3.The emotion recognition based on the improved Bi-ProgNet further proves the effectiveness of the lifelong learning-based emotion recognition method.On the basis of previous experiments,in order to ensure the fairness of the experiment and the universality of the method proposed in this thesis,the network structure of Bi-ProgNet is optimized based on the public datasets.The improved Bi-ProgNet adds a spatial attention mechanism module in the process of knowledge transfer,selects the transfer knowledge of face emotion recognition network through network adaptive learning,gives important knowledge a higher weight,and reduces useless and redundant information.proportion,making the knowledge transfer between networks more accurate.The optimized model can achieve 69.94% and 75.56% accuracy on the Fer2013 face dataset and SEED EEG dataset,respectively.Although the prediction accuracy is not optimal compared to the models that perform emotion recognition on these two datasets respectively,this model achieves emotion recognition for multiple tasks,and maintains a high accuracy in both tasks at the same time.