Emotion Recognition Based on Dynamic Energy Features Using a Bi-LSTM Network.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models for electroencephalogram (EEG) signal emotion recognition often struggle with low-resolution data and small sample sizes.
• Existing methods face challenges in extracting high-quality features from complex EEG signals, limiting model performance.
• Noise superposition during feature analysis can further degrade the effectiveness of EEG-based emotion recognition systems.

Purpose of the Study:

• To propose a novel deep network model for EEG emotion recognition that overcomes limitations of low resolution and small datasets.
• To introduce dynamic energy features for more robust and informative representation of EEG signals.
• To enhance the applicability of deep learning models in scenarios with limited EEG data.

Main Methods:

• Proposed a deep network model leveraging dynamic energy features extracted from EEG signals.
• Introduced the concept of an energy sequence to mitigate noise superposition during feature extraction.
• Utilized fully connected layers and bidirectional long short-term memory (Bi-LSTM) networks to accommodate small datasets.

Main Results:

• Achieved high accuracy rates on benchmark datasets: 89.42% on SEED and 77.34% on DEAP.
• Demonstrated the effectiveness of dynamic energy features in capturing temporal persistence and multicomponent complexity of EEG signals.
• Validated the model's robustness using leave-one-subject-out (LOSO) and 10-fold cross-validation (CV) strategies.

Conclusions:

• The proposed deep network model based on dynamic energy features significantly improves EEG emotion recognition accuracy.
• The method is particularly effective in handling small datasets, a common challenge in brain-computer interface research.
• Dynamic energy features offer a promising approach for advanced analysis of EEG signals in emotion recognition tasks.