Spectrogram-driven Emotion Detection from Electroencephalogram

Emotion detection aims to interpret emotions through data like text, voice, and physiological signals, which holds significant potential for monitoring mental health and human-computer interaction. Electroencephalogram (EEG) is a non-invasive technique recording brain activity and offers unique insights for real-time emotion detection. This study evaluates classifying emotions (i.e. positive, negative, neutral) in EEG using deep learning, emphasizing Convolutional Neural Networks (CNN). After going through the preprocessing part, the time-frequency representations along with random transformations were created as inputs to the CNN architecture by a short-time Fourier transform of the EEG signals from the SEED dataset. The spectrograms have been processed to have better outcomes, and they showed an overall accuracy of 99.80% in detecting the emotions classes. Also, processed spectrograms, along with the actual spectrograms, have been fed together to the CNN model as a hypothesis, and there we achieved an overall accuracy of 66%, depicting that the usage of processed spectrograms can be a potential field of study to recognize emotions from EEG signals.