Emotion Recognition From EEG Based On Generative Adversarial Networks

Emotion is an important part of human life.Emotion recognition is of great use for fields like human-computer interface and treatments of emotional disorders.Emotion recognition can be completed through two approaches,namely,non-physiological signals(such as facial expressions,voices and behaviors)and physiological signals(such as ECG and EEG).Emotion recognition based on non-physiological signals is hard to be solid as the signals can be faked.Physiological-signal based emotion recognition is not easy to fake but is facing another challenge:the time-varying nature of physiological signals.In this paper,we concentrated on the time-varying nature of EEG signals,and proposed a new EEG-emotion recognition model to address this problem.The contribution includes:1.Analyzed the time-varying nature of EEG signals through multiple experiments.It shows that EEG signals from one subject within a short period of time show good commonality,where signals from different emotions show good divisibility.But signals at different time periods have poor divisibility,as most differences between signals are caused by time-varying noise and unpredictable changes in the environment.Therefore,common machine learning algorithms have poor performance on such EEG-based emotion recognition jobs.2.Proposed a new emotion recognition model based on Generative Adversarial Networks(GANs)and the Wasserstein distance.The model comes with two training approaches,namely,domain adversary and domain adaptation.The domain adaptation approach of the proposed model achieved the accuracy of 86.48%on subject-dependent cross-session emotion classification experiments,which outperforms other baseline method and state-of-the-art models.Correlation between different EEG frequency bands is also tested in experiments,where results show that Beta and Gamma frequency bands are most effective in emotion recognition.3.Designed and developed an online EEG-emotion recognition system,including data acquisition,real-time signal visualization,model training and real-time emotion recognition functions.This system satisfies the need of online EEG-based emotion recognition experiments and provides a solid platform for related researches in future.In all,the model and the system developed in this paper are able to complete EEG-based emotion recognition tasks,and is worth further research.