Emotion Recognition Using Few-Shot Learning with Embedding Space and Prototype-Based EEG Analysis

EEG-based emotion recognition is a promising area of research that uses brain patterns of activity to determine the human emotional state. In this research, multiclass emotion detection using DEAP dataset and classifying six diverse emotions, including happiness, sadness, anger, surprise, calm, and fear, is demonstrated. A novel preprocessing technique called Dynamic Adaptive Signal Denoising and Normalization is proposed in order to improve the quality of EEG signals by removal of noise artifacts and normalization of amplitudes of the signal. To further enhance model performance, the Emotion-Specific Synthetic EEG Generation Framework is used to synthesize more samples for each emotion category in data augmentation. Feature extraction was performed using the Multi-Domain Emotion Feature Extraction Framework, where information was drawn from five domains: time, frequency, time-frequency, spatial, and temporal. It then identifies the most relevant features through the Adaptive Emotion-Specific Feature Selection (AEFS) method in order to optimize the dataset for accurate classification. The EmotionSpecific Few-Shot Learning (ES-FSL) is used to achieve robust classification even with a few labelled samples. The proposed approach attains a significant accuracy of 92.5%, outperforming the traditional emotion recognition methods. The proposed comprehensive methodology, from advanced preprocessing to data augmentation techniques, up to feature extraction and selection techniques, provides a robust framework for emotion recognition from EEG signals. This work advances state-of-theart conditions in emotion recognition and ensures scalability along with adaptation in real-world applications where possibly labelled data is limited. The proposed method stands as a significant step toward making emotion recognition systems more accurate and applicable.