A performance analysis of deep learning algorithms for emotion detection based on EEG

Human emotions are a person's reaction to events around him. Human emotions play an important role in a person's life and greatly impact his behavior based on his daily mood. The development of computers in the past two decades has led to increased interest in human-machine interaction applications. Among these applications are human emotions applications, which have gained a lot of attention lately by psychologists. The main aim of this thesis is to detect human emotions using deep learning techniques. In this thesis, a physiological data set GAMEEMO was used to classify human emotions based on the discrete emotions model and the dimensional emotion model. The GAMEEMO dataset is an EEG dataset consisting of 128 Hz signals for 28 subjects. There are three main tasks of this thesis. The first task: Four feature extraction methods (statistical features (SF), a combination of SF and Welsh power spectral density (PSD), a combination of SF and Fast Fourier transform (FFT), and the combination of SF and wavelet packet decay (WPD) are applied on the pre-processed EEG signals to extract more features. The second task: Split the data based on two classes and four classes. The third task: Build a convolutional neural network (CNN) using a 1D CNN, a recurrent neural network (RNN) using a long-term memory (LSTM), and a hybrid model 1D CNN + LSTM architectures. This thesis achieved high performance using the WPD+SF feature extraction method, with more than 98% accuracy in both emotions' models.