STRAViT: A Spatial-Temporal Feature-Reshaped Agent Vision Transformer for EEG-Based Emotion Recognition

Emotion recognition from electroencephalogram (EEG) signals has attracted increasing interest due to its applications in affective computing and brain-computer interaction. However, effectively integrating spatial-temporal features while attending to emotionally relevant brain regions remains a significant challenge. In this article, we propose a spatial-temporal feature-reshaped agent vision transformer (STRAViT) network, which fuses differential entropy (DE) and functional connectivity matrices through a dual-branch attention-based feature reshaping module (FRM), enabling refined feature integration. In addition, we introduce an agent vision transformer (AViT) that utilizes learnable agent tokens to capture global-local dependencies within EEG representations efficiently. Extensive experiments conducted on the SEED and SEED-IV datasets demonstrate that the STRAViT achieves classification accuracies of 98.7% and 87.2% on SEED, 95.5% and 72.5% on SEED-IV, for discrete emotion recognition, under subject-dependent and subject-independent strategies. The proposed network not only effectively integrates spatial-temporal features of EEG signals but also enhances the modeling of inter-regional dependencies through agent attention. Comprehensive tests confirm that the STRAViT yields excellent performance on emotion recognition tasks.