Fusion of Convolutional Neural Networks for P300 based Character Recognition

Abstract

In this work, a brain-computer interface (BCI) system for character recognition is proposed based on P300 signal. Signal classification is the most challenging task in electroencephalography (EEG) signal processing as the amplitude of the EEG signal is low and it can be affected by the surrounding noise. The feature extraction is an important step for any classification task. The manually designed features are not sufficient to represent the signal properly due to the subject and surrounding environment variability. There is a need to develop feature extraction techniques which will extract highlevel feature from the raw data automatically. In this work, two parallel CNN model with different kernel size is proposed to extract multi-resolution feature from the dataset. The proposed CNN model extracts spatial and temporal feature from the dataset. To mitigate the over-fitting problem, dropout is used before the fully-connected layer of the CNN architecture, which improves the network performance. The scores of these two CNN architectures are fused together for P300 detection. Also, ensemble of CNN (ECNN) architecture is proposed to reduce the variation between the classifiers and enhance the character recognition performance. The proposed method is tested on BCI Competition III dataset and the results are fairly comparable and better with the earlier methods.