Electroencephalograph based Human Emotion Recognition Using Artificial Neural Network and Principal Component Analysis

In recent decades, automatic human emotion detection plays a crucial role in human and machine interaction. Electroencephalograph (EEG) based human emotion detection is a challenging process due to the diversity, and complexity of human emotions. For recognizing diverse emotions, a novel model is presented in this paper. Initially, an average mean reference technique is used to eliminate the environmental artifacts, instrumentation artifacts, and biological artifacts from the EEG signals, which are collected from DEAP dataset. Next, feature extraction is carried out using Fast Fourier transform (FFT) with Power Spectral Density (PSD) to extract feature vectors from the denoised EEG signals. Further, feature dimensionality reduction is performed utilizing Principal Component Analysis (PCA) to diminish the dimensions of the extracted features. A total of 230 EEG feature vectors are given as the input to Artificial Neural Network (ANN) for classifying valence and arousal emotion states. The proposed PCA-ANN model performance is validated in terms of average classification accuracy and f-score. The experimental outcome demonstrates that the proposed PCA-ANN model achieved an improved accuracy in emotion classification, which is effective compared to the existing models such as ensemble learning algorithm, a convolutional neural network with the statistical method, and sparse autoencoder with logistic regression. The proposed PCA-ANN model achieved 87.14% and 86.31% of accuracy in valence and arousal states, and obtained 90.45% and 92.03% of f-score value in valence and arousal emotion states.