Abstract
The recognition of human emotions based on the analysis of non-stationary
electroencephalogram (EEG) signals is important to design a robust human–
computer interaction (HCI) model. This contribution investigates the feasibility of
using the Stockwell transform (ST) based time–frequency (T–F) analysis of nonstationary EEG signals for emotion recognition. To improve the resolution of
EEG signals in the T–F plane, the traditional Gaussian window present in the ST
was replaced by an adaptive signal-dependent modified window whose parameters are optimally selected using maximum energy concentration measure-based constrained optimization. The T–F images of the processed EEG signals obtained by employing the proposed modified ST were subsequently fed to a pretrained, highly dense convolutional neural network (CNN) in order to classify different emotional states. In this work, both subject-specific as well as cross-subject-based classification was performed to test the robustness of the proposed technique. The proposed methodology was verified on two publicly available databases, named the SJTU Emotion EEG Dataset (SEED) and SEED-IV. It was observed that the highest mean accuracies of 98.6% and 89.9% for the SEED and the SEED-IV databases, respectively, were obtained considering subject-specific classification. In the case of cross-subject classification, the proposed method delivered an average accuracy of 93.26% for the SEED and 78.6% for the SEED-IV database, respectively. We observed that the performance of the proposed method was found to be significantly better than the existing literature studies on the same datasets, which validates the efficiency of the proposed method to develop an efficient human–computer machine interface system.
electroencephalogram (EEG) signals is important to design a robust human–
computer interaction (HCI) model. This contribution investigates the feasibility of
using the Stockwell transform (ST) based time–frequency (T–F) analysis of nonstationary EEG signals for emotion recognition. To improve the resolution of
EEG signals in the T–F plane, the traditional Gaussian window present in the ST
was replaced by an adaptive signal-dependent modified window whose parameters are optimally selected using maximum energy concentration measure-based constrained optimization. The T–F images of the processed EEG signals obtained by employing the proposed modified ST were subsequently fed to a pretrained, highly dense convolutional neural network (CNN) in order to classify different emotional states. In this work, both subject-specific as well as cross-subject-based classification was performed to test the robustness of the proposed technique. The proposed methodology was verified on two publicly available databases, named the SJTU Emotion EEG Dataset (SEED) and SEED-IV. It was observed that the highest mean accuracies of 98.6% and 89.9% for the SEED and the SEED-IV databases, respectively, were obtained considering subject-specific classification. In the case of cross-subject classification, the proposed method delivered an average accuracy of 93.26% for the SEED and 78.6% for the SEED-IV database, respectively. We observed that the performance of the proposed method was found to be significantly better than the existing literature studies on the same datasets, which validates the efficiency of the proposed method to develop an efficient human–computer machine interface system.
Original language | English |
---|---|
Title of host publication | Modelling and Analysis of Active Biopotential Signals in Healthcare |
Publisher | IOP Publishing Ltd. |
Chapter | 2 |
Pages | 1 - 28 |
Number of pages | 28 |
Volume | 2 |
ISBN (Electronic) | 978-0-7503-3411-2, 978-0-7503-3410-5 |
ISBN (Print) | 978-0-7503-3409-9, 978-0-7503-3412-9 |
DOIs | |
Publication status | Published (in print/issue) - 1 Dec 2020 |
Keywords
- EEG
- Emotion Detection
- Deep Leraning
- CNN
- DenseNet