EEG based Emotion Recognition

Abstract

Accurate detection of emotion plays a vital role in field of Human-Robot Interaction (HRI). Human Robot Interaction (HRI) robots aim for social intelligence, allowing them to recognize and respond to human emotional state. This research study, proposed a state-of-the-art ap proach Model soup for the emotion recognition task using EEG signals into their respective classes. Model soup generates a model by averaging the weights of fine-tune model which reduces time and model complexity. EEG signals are converted to scalograms by Continuous wavelet transform such time-frequency representation corresponds effectively to time-varying nature of EEG patterns, followed by normalization and data augmentation. The "Model Soup" is used to average the weights of these three fine-tuned GoogLeNet, ResNet-34, and VGG-16 models to improve accuracy without requiring additional inference time and resources. Finally, Grad-Cam visualization is employed to improve interpretability and then model is evaluated on SEED dataset, achieves a 99.31% recognition accuracy. Using model soup improves the emotion recognition accuracy and out performs other state-of-the-art deep learning models.