In recent years, EEG emotion recognition has had broad prospects in various research fields. Most of the existing EEG identification studies use a single feature and do not simultaneously consider the effects of EEG information on time, space and frequency. However, neuroscience shows that the response degree of emotion depends on the EEG channels of different frequencies. In this paper, A recursive neural network approach is proposed to analyze EEG signals at different times and regions by integrating the time domain, frequency domain, and space domain. Firstly, the 2D channel map is extracted to obtain frequency domain features, and then the 2D channel map with different frequency domain features is fused to obtain 3D spatial frequency domain data. Finally, the 3D data in different time periods are connected to obtain the 4D feature structure used for the training depth model. 4D data feature makes up for the defect that a single feature cannot perceive channel information of different frequencies and produce a more significant effect than single feature emotion recognition. Experiments on the SEED data set show that the multi-feature fusion model performs better than other models in EEG emotion recognition.
4D Recurrent Neural Network Based on Time-Space-Frequency Domain Fusion for EEG Emotion Recognition
2022-10-12
1210704 byte
Conference paper
Electronic Resource
English
Abnormal Emotion Recognition Based on Audio-Visual Modality Fusion
Springer Verlag | 2023
|