Loading...

EEG-based Emotion Recognition Using Graph Learning

Talaie, Sharareh | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57131 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Hajipour Sardouie, Sepideh
  7. Abstract:
  8. The field of emotion recognition is a growing area with multiple interdisciplinary applications, and processing and analyzing electroencephalogram signals (EEG) is one of its standard methods. In most articles, emotional elicitation methods for EEG signal recording involve visual-auditory stimulation; however, the use of virtual reality methods for recording signals with more realistic information is suggested. Therefore, in the present study, the VREED dataset, whose emotional elicitation is virtual reality, has been used to classify positive and negative emotions. The best classification accuracy in the VREED dataset article is 73.77% ± 2.01, achieved by combining features of relative power (RP) in the four frequency bands of theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz), and gamma (30-49 Hz) using an SVM classifier. Among the frequency bands, the best classification belongs to the theta frequency band with an accuracy of 71.35% ± 2.01, utilizing RP features and an SVM classifier. In this research, initially, using RP features in the theta frequency band and a non-linear function for the SVM classifier kernel, the classification accuracy increased to 74.17% ± 5.59. Adding a sequential feature selection method improved the classification accuracy to 76.28% ± 2.61. Furthermore, graphs were utilized to understand hidden structures and relationships between brain regions. By extracting graph features from learned graphs based on RP, the classification accuracy enhanced to 79.47% ± 3.15. Subsequently, the positive impact of common spatial pattern (CSP) features was demonstrated, increasing classification accuracy to 82.14% ± 1.44. In this study, two graph-based CSP algorithms were proposed. The first algorithm achieved a classification accuracy of 83.04% ± 1.96 using known graphs based on RP features, and the second algorithm reached a classification accuracy of 80.55% ± 2.52 using unknown graphs learned in a recursive learning algorithm. An important point to note is the difference between the learned graph in the recursive algorithm and the RP-based learned graph, indicating that in the face of negative emotional states, the frontal lobe will be significantly involved, similar to the occipital lobe
  9. Keywords:
  10. Feature Extraction ; Common Spatical Patterns ; Emotion Recognition ; Electroencephalogram (EEG)Emotion Recognition ; Graph Learning ; Electroencphalogram Signal ; Relative Power

 Digital Object List

 Bookmark

No TOC