Fine-Grained Emotion Recognition from EEG Signal Using Fast Fourier Transformation and CNN

Emotions are mental states originating in the human brain, and this is closely related to the activities of the nervous system. Electroencephalogram (EEG) is a well-established approach to record neuron activities which is reliable for emotion recognition compared to the non-physiological clues. So far, there have been reports of various researches searching for active patterns involving different emotions. However, most of the previously published system could only classify 4 human emotions using the technique of binary classification but humans have more and complex emotions which couldn't be captured with only 4 classes. So, we proposed a fine-grained emotion classification technique which can classify 64 emotions including all the complex emotions. Hence, this paper presents convolutional neural network (CNN) models working on the DEAP dataset, and it contains emotional states which are arousal, valence, dominance and liking. Our binary models achieved 96.63% and 96.18% accuracy respectively for valence and arousal. Only four emotions are found with binary classification whereas 8-class classification can precisely recognize 64 emotions. The 8-class classification achieves a promising accuracy of 93.83 % and 93.79% respectively for valence and arousal. For both cases, Fast Fourier Transformation (FFT) has been used as the feature extraction method and all the four classification models are created under 1D-CNN using the same architecture.