Deep Ensemble Learning Approach for Multimodal Emotion Recognition

In this research initiative, our central objective is the development of an emotion recognition system through the utilization of cutting-edge deep learning methodologies. Our methodology involves adopting a subject-independent approach to classify emotions based on Electroencephalogram (EEG) signals sourced from a standard benchmark DEAP dataset. To ensure data quality and reliability, we initiate our process by meticulously preprocessing the raw signal data, which includes the application of Normalization and Common Average Reference (CAR) techniques. Subsequently, we employ Discrete Wavelet Transform (DWT) technique to extract salient features from the cleaned EEG data. These extracted features serve as the foundation for training three distinct deep learning models: the CNN-LSTM, CNN-GRU, and 2D-CNN models. To consolidate their predictive capabilities, we employ a Majority voting algorithm, effectively combining the strengths of these models. Notably, our proposed deep ensemble learning approach yields an impressive accuracy rate of 88% when evaluated on the challenging DEAP dataset.