An EEG Data Processing Approach for Emotion Recognition
Emotion recognition is a significant area of research in affective computing. The emotion recognition methods based on facial expressions can be deceptive due to fake expressions, poor or changing light, or other environmental conditions. With the technological advancements in machine learning and deep learning, EEG-based emotion recognition has attracted interest from scientists.
This paper proposes a novel emotion recognition approach that depends on a reduced number of EEG electrode channels and a normalization method that can overcome the negative impact of individual differences to attain high identification accuracy.
The power spectral density (PSD) is a commonly used and well-accepted feature in the analysis of human EEG activities. The Kruskal-Wallis test is applied to statistically examine the difference in PSD features among different emotions to determine the emotion-sensitive channels.
Six prospective sets of EEG electrode channels are identified based on the statistical testing results of PSD characteristics taken from the SJTU Emotion EEG Dataset (SEED). PSD features and differential entropy (DE) features of each set are delivered to six classical classifiers used for emotion recognition, including K Nearest Neighbors (KNN), Logistic Regression (LR), Support Vector Machine (SVM), Gaussian Naive Bayes (GNB), Decision Tree(DT), Random Forest(RF), eXtreme Gradient Boosting(XGB), Multilayer Perceptron(MLP), Bootstrap Aggregating(BA), Convolutional Neural Network(CNN) and Long Short Term Memory Network(LSTM) to assess the effectiveness of various channel sets on emotion recognition.
The results demonstrate that when the number of channels is lowered, the emotion recognition accuracy rises and becomes comparatively steady. The best set is only nine channels, mainly from the temporal lobes. Topo maps of the PSD features across the three emotions from different subjects are performed to investigate further the relationship between the selected channels and human emotion.
The results show that the activated channels with human emotions mainly distribute in the temporal lobe of brain regions and prove individual differences across subjects in each emotion.
Batch normalization (BN), which aims to reduce the impact of individual differences on emotion recognition, is introduced. After subjecting the candidate sets to batch normalization (BN), the normalized features are utilized to identify the emotions with the classifiers. The experimental findings showed that the recognition accuracy achieved using a subset of the available electrodes is almost the same or even better than the results obtained using all the channels.
The simple operations of BN make it excellently adaptable to various algorithms, including machine learning and deep learning. The study demonstrates that applying BN improves each channel's significance level with a statistically increasing number of channels among the emotions under study. Using batch normalization features from fewer channels can lead to easier and better emotion detection performance.
EEG-based emotion recognition is significant in the passive brain-computer interface (BCI). It aims to improve communication, develop strategies to regulate negative emotions, and create new applications involving human and machine interfaces or interactions.
The proposed method can be applied to collect EEG signals and process them with less computational cost and enhanced performance on emotion recognition. Though it has limitations, it has created opportunities for future research and development.