The Dual-3DCRU Model Based on Hemisphere Asymmetry for EEG Emotion Recognition

In this paper, we propose an innovative multi scale model called Dual-3DCRU, designed to extract more discriminative feature from both the left and right hemispheres. Byleveraging spatial relationships of electrode locations and the distinct frequency bands present in EEG signals from brain two hemispheres regions, we reconstructed comprehensive three-dimensional frequency space fusion feature representation. The Dual-3DCRU utilizes two different multi scale 3D -CNN to effectively capture both frequency and spatial information embedded in the 3D feature representations. Furthermore, we integrate a Gated Recursive Unit (GRU) to model the temporal dynamics across continuous time periods and extract critical temporal information. Ultimately, the Dual-3DGRU model delivers a holistic feature representation that incorporates time, frequency, and spatial information seamlessly. Our experiments showcase that the Dual-3DCRU model attains exceptional performance on DEAP dataset. Specifically, the accuracy in the Valence dimensionis 94.35%, F1-score is 93.24%, accuracy in the Arousal dimensionis 95.04%, and F1-score is 94.17%.