Understanding the visual-evoked eeg image reconstruction and classification of digits and alphabets | |
| Author | Jakrapop Akaranee |
| Call Number | AIT Thesis no.CS-22-02 |
| Subject(s) | Electroencephalography Diagnostic imaging--Data processing Image Processing, Computer-Assisted |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Computer Science |
| Publisher | Asian Institute of Technology |
| Abstract | Recovery from what we have seen from a recording of electroencephalograph (EEG) or EEG image reconstruction has been a hot topic in brain-computer interface (BCI). However, there is recent controversy where studies could not reproduce the accuracy claimed. This issue arises from the experiment design and preprocessing used in the original work. This study aims to investigate the controversy on our custom dataset. We collect three dif ferent datasets where the subjects viewed image of digits or alphabet with different EEG experiment setup. Then, we analyze the factors/components of EEG image reconstruction models and EEG classification for better understanding. Our findings support previous findings in R. Li et al. (2020); Ahmed, Wilbur, Bharadwaj, and Siskind (2021) where they suggested that Kavasidis, Palazzo, Spampinato, Giordano, and Shah (2017) was able to achieve satisfying accuracy due to many flaws in their method. First, temporal correlation should be kept in consideration when designing EEG experiments since a longer experiment time may cause temporal correlation. Anyhow, following a rapid event design can avoid the effect of temporal correlation. Second, we demonstrated that data preprocessing in X. Zhang et al. (2020) where the data were segmented first before spitting it into training and test set, could cause contamination between train and test sets, resulting in misleading accuracy. Lastly, combining the data of all subjects and then training a subject-independent model is unsuitable for EEG image reconstruction task since no study has confirmed if the data can be combined across subjects. We further examined with the stimuli display time, EEG features, and models to better under stand EEG classification, which is the foundation of EEG reconstruction. Our result suggest that stimuli display time of 0.5 is enough for EEG classification, an alpha band (8-12 Hz) is the best feature for classification, and EEGNet performs better than eleven other machine learning and deep learning models for EEG classification. The contribution of our study is to emphasize the challenge of the EEG image reconstruction task and raises awareness of rigorous and careful experiments, especially in delicate data such as EEG. |
| Year | 2022 |
| Type | Thesis |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Computer Science (CS) |
| Chairperson(s) | Chaklam Silpasuwanchai |
| Examination Committee(s) | Dailey, Matthew N.;Attaphongse Taparugssanagorn |
| Scholarship Donor(s) | National Science and Technology Development Agency (NSTDA), Thailand |
| Degree | Thesis (M. Sc.) - Asian Institute of Technology, 2022 |