EEG-Based Emotion Recognition Fusing Spacial-Frequency Domain Features and Data-Driven Spectrogram-Like Features

Research on emotion recognition based on EEG (electroencephalogram) signals has gradually become a hot spot in the field of artificial intelligence applications. The recognition methods mainly include designing traditional hand-extracted features in machine learning and fully automatic extraction of EEG features in deep learning. However, onefold features cannot represent emotional information perfectly which is contained in EEG signals. Traditional hand-extracted features may lose a lot of hidden information contained in raw signals, and automatically extracted features also do not contain prior knowledge. In this context, a multi-input Y-shape EEG-based emotion recognition neural network is proposed in this paper, which fusing spacial-frequency domain features and data-driven spectrogram-like features. It can effectually extract information in three domains, time, space, and frequency from raw EEG signals. Moreover, this paper also proposes a novel EEG feature mapping method. The experimental results show that the accuracy of EEG emotion recognition has achieved the state-of-the-art result based on the established DEAP benchmark dataset. The average emotion recognition rates are 71.25%, 71.33% and 71.1% in valance, arousal and dominance respectively.