A Self-Attention-Based Trans-Generative Adversarial Network with Relative Position Encoding for Emotion Recognition Using EEG Signals
Abstract
Emotions are essential components of human activities, affecting cognitive processes such as decision-making, planning, and reasoning, while also shaping social interactions and behavioral responses. Recognition of emotions can be achieved through diverse sources, including facial expressions, speech patterns, bodily gestures or posture, and physiological signals. Physiological signals provide a valuable means of enhancing the objectivity and reliability of emotion detection. Among these, the Electroencephalogram (EEG) recordings hold particular importance, as it originates directly from the central nervous system and exhibit a strong correlation with emotional states. EEG signals offer high spatial resolution, enabling detailed evaluation of brain activity and making them a widely adopted modality in emotion recognition research. It faces challenges such as variability introduced by electrode placement, differences in signal morphology across individuals, and the complexity of selecting appropriate features from EEG data. Advances in artificial intelligence, encompassing traditional machine learning and modern deep learning approaches, have demonstrated superior effectiveness in managing the complexity and diversity of EEG signals linked to emotional states. Thus, this paper focuses on proposing a robust emotion recognition model by incorporating the strength of deep learning methodology. At the processing phase, the essential EEG signals are collected from the EEG DEAP dataset. Then, the gathered signals are converted into spectrogram images using Short-Time Fourier Transform (STFT). Later, the emotion recognition procedure is executed on the designed Self-Attention-based Trans-Generative Adversarial Network with Relative Position Encoding (SA-TGAN-RPE) framework using the resultant spectrogram images. Finally, various validations are carried out to showcase the efficacy of the designed network over conventional techniques.