Multi-Region and Multi-Band Electroencephalogram Emotion Recognition Based on Self-Attention and Capsule Network

• DOI: 10.3390/app14020702
• Publication Date: 2024-01-15T13:52:03Z
• Authors: Sheng Ke; Chaoran Ma; Wenjie Li; Jidong Lv; Ling Zou
• Keywords: Radio spectrum; Frequency band

Research on emotion recognition based electroencephalogram (EEG) signals is important for human detection and improvements in mental health. However, the importance of EEG from different brain regions frequency bands different. For this problem, paper proposes Capsule–Transformer method multi-region multi-band recognition. First, features are extracted combined into feature vectors which input fully connected network dimension alignment. Then, inputted Transformer calculating self-attention among to obtain contextual information. Finally, utilizing capsule networks captures intrinsic relationship between local global features. It merges bands, adaptively computing weights each region band. Based DEAP dataset, experiments show that achieves average classification accuracies 96.75%, 96.88%, 96.25% valence, arousal, dominance dimensions, respectively. Furthermore, conducted individual or it was observed frontal lobe exhibits highest accuracy, followed by parietal, temporal, occipital lobes. Additionally, performance superior high-frequency band compared low-frequency signals.