EEG based emotion recognition using wavelets and neural networks classifier

Abstract

Emotions have a vital role in the day-to-day life of human beings, the need and importance of emotion recognition systems have increased with the role of human computer interface applications. In this paper, machine learning methods are used to model a relationship using the publicly available dataset SEED (SJTU Emotion EEG Dataset) which contains EEG signals of 15 participants recorded when excited to video stimuli. The signal processing techniques in time domain and time-frequency domain (Wavelet analysis) are used to extract the desired features. The discrete wavelet transforms are used to extract frequency bands. The features such as Statistical features, Hjorth parameters, differential entropy, and the combination on symmetric electrodes (differential asymmetry DASM and rational asymmetry RASM) are extracted. Artificial neural networks and Support Vector Machine (SVM) are applied on the feature set to develop prediction models to extract the emotion information carried by the participant from emotional characteristics exhibited in different frequency bands. These models are evaluated on the dataset and emotions are classified using ANN into three different states such as positive, negative and neutral states with an accuracy of 91.2%.