Emotion Detection using Stack Auto Encoder, Deep Learning and LSTM Model

Abstract

There is an increasing need for distress emotion acknowledgement and channeling the same is very important. There is an increasing need for machines to understand human and their complex emotions deeply. This research describes a unique framework for emotion detection that helps brain-computer interface/machine (BCI) to understand human emotions and brain complexity and working using multi-channel electroencephalograms (EEG). We have employed two datasets in our work, Database for Emotion Recognition Through EEG and ECG Signals from Wireless Low-Cost Off-the-Shelf Devices (DREAMER) and dataset for emotion analysis using EEG, physiological and video signals (DEAP) for LSTM model implementation and validation respectively. A linear model is considered for EEG signal mixing and an emotion timing model comprise the framework. Leveraging the contextual correlations within EEG feature sequences, our proposed methodology effectively recovers EEG source signals from the obtained EEG signals, thereby improving the accuracy rate of classification. Also, stress bins were set up for individual users to assess their degree of stress and calmness following exposure to external stimuli. DEAP dataset using LSTM framework was enforced for emotion recognition, and mean recognition accuracy using area under curve as evaluation matrix for valance and arousal was 82.02% and 76.52% respectively, validating competence of framework. Novelty of our work is it improved competency in feature extraction, use of context correlations increasing accuracy and use of spatio-temporal features in the proposed model framework.