Please use this identifier to cite or link to this item:
https://www.um.edu.mt/library/oar/handle/123456789/92091
Title: | Real-time EEG-emotion recognition using prosumer grade devices |
Authors: | Borg Bonello, Francesco (2021) |
Keywords: | Electroencephalography Machine learning Real-time data processing Neural networks (Computer science) Support vector machines Emotions -- Computer simulation Human-computer interaction |
Issue Date: | 2021 |
Citation: | Borg Bonello, F. (2021). Real-time EEG-emotion recognition using prosumer grade devices (Bachelor’s dissertation). |
Abstract: | Electroencephalography-based Emotion Recognition (EEG-ER) is a widely researched technique that allows the detection of emotions based on one’s brain signals. Machine learning solutions consider data collected from high-end devices, thus providing high-dimensional data to classify emotions based on brain signals. Although recent years have seen the launch of lower-costing EEG products, there has been a lack of attention given to classifying real-time data from these low-end devices that consist of a reduced number of channel data. In this study, we build models based on both subject-independent as well as subject-dependent data that classify Valence and Arousal dimensions which in turn locate an emotion based on Russell’s Circumplex Model of Affect. We first devise solutions to conduct real-time EEG-ER using data from a high number of channels, which include 3DCNN as well as SVM. We then apply these models to a reduced-channel version of the DEAP dataset which consists of only 5 channels. A comparison is made between high-end and low-end solutions, ultimately determining the viability of low-end EEG-ER. Results show that using the baseline removal preprocessing technique reports an enhanced overall real-time classification accuracy for both the full-channel (32 channel) data as well as the reduced-channel (5 channel) datasets. Our full-channel SVM model achieves state-of-the-art subject-dependent accuracy with 95.3% and 95.7% on the Valence and Arousal dimensions, with the reduced-channel solution only decreasing in accuracy by 3.46% and 3.71%. This slight decrease is an encouraging result due to the fact that even though a reduced number of channels are being considered, the high standard set by the full-channel model is retained. |
Description: | B.Sc. IT (Hons)(Melit.) |
URI: | https://www.um.edu.mt/library/oar/handle/123456789/92091 |
Appears in Collections: | Dissertations - FacICT - 2021 Dissertations - FacICTAI - 2021 |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
21BITAI015.pdf Restricted Access | 2.54 MB | Adobe PDF | View/Open Request a copy |
Items in OAR@UM are protected by copyright, with all rights reserved, unless otherwise indicated.