Hybrid CNN-transformer architecture for enhanced EEG-based emotion recognition: capturing local and global dependencies with self-attention mechanisms

Conventional EEG-based modeling structures struggle to effectively process the combination of spatial information and temporal data. It can detect spatial patterns but cannot track extended neural dependencies. On the other hand, transformers understand global patterns well but may overlook fine-grained local relationships. In this paper, presents a hybrid CNN-transformer architecture that combines the strengths of both data, specially designed for emotion recognition from EEG signals. This architecture improves CNNs’ spatial pattern detection capabilities and transformers’ self-attention mechanism for global pattern understanding. The performance of the proposed architecture is evaluated using the DEAP dataset, comprising 40 EEG sessions collected from 32 subjects. Results convey that 87% accuracy on the DEAP dataset which is outperformed than other models include AlexNet (83.50%), VGG-16 (85.00%), ResNet-50 (85.50%), GoogleNet (85.00%), and MobileNetV2 (86.00%). Also, maintains high performance regardless of changing emotion intensity levels and demonstrates robustness for practical applications.