Brain-Computer Interface

10Identification of Imagined Bengali Vowels from EEG Signals Using Activity Map and Convolutional Neural Network

Rajdeep Ghosh¹*, Nidul Sinha² and Souvik Phadikar²

¹School of Computing Science and Engineering, VIT Bhopal University, Kotri Kalan, Madhya Pradesh, India

²Department of E.E., NIT Silchar, Silchar, Assam, India

Abstract

Classification of electroencephalogram (EEG) signals is challenging due to its non-stationary nature and poor spectral resolution w.r.t. time. This chapter presents a novel feature representation termed as activity map (AM) for the classification of imagined vowels from EEG. The proposed AM provides a visualization of the temporal and spectral information of the EEG data. The AM is created for each of the EEG recorded from 22 subjects imagining five Bengali vowels /আ/, /ই/, /উ/, /এ/, and /ও/ using 64 channels. The AM is generated by extracting the band power in delta, theta, alpha, beta, and gamma bands for each second of the EEG and subsequently stacking them to form a matrix. The matrix is then converted into a heat map of 100 × 200 pixels called AM. The AM thus obtained, is classified using a convolutional neural network (CNN), achieving an average accuracy of 68.9% in classifying the imagined vowels. The CNN demonstrates superior performance in comparison to other methods reported in the literature using various features such as common spatial pattern (CSP), discrete wavelet transform (DWT), etc. and with different classifiers such as kNN (k nearest ...