AI-BASED EEG SIGNAL PROCESSING FOR EMOTION RECOGNITION AND MIND STATE ANALYSIS

Abstract

In human-computer interaction (HCI), emotion detection is essential because it allows device to interpret and react to human emotional states by using neurophysiological data. Electroencephalogram (EEG) signals are one of the most common signals used for classifying emotions as they provide online and non-invasive assessment of brain activity. This thesis focuses on the comparative analysis between standalone deep learning and hybrid deep learning models for emotion recognition using EEG signals. A detailed evaluation is conducted on various classification techniques applied to the SEED-IV dataset. Based on the performance of the models, a hybrid classifier that combines long short-term memory (LSTM) networks to capture temporal relationships, convolutional neural networks (CNN) to extract spatial characteristics, and a convolutional block attention module (CBAM) to increase attention to relevant spatio-temporal aspects in the EEG spectrogram, hereafter referred as CNN-LSTM-CBAM, is proposed. The proposed CNN-LSTM-CBAM architecture achieves improved classification accuracy of 68.67% on the test set and emotion-wise consistency in emotion recognition tasks compared to the standalone models. The results validate the possibility of incorporating different deep learning concepts and attention processes into hybrid deep learning pipelines for efficient EEG data processing and interpretation in applications related to affective computing.