Abstract:

In the literature, it is encountered a vast amount of studies related to the production of controllable wheelchairs for people with disabilities or the prediction of activity thought to be performed. In general, the electroencephalography (EEG) signal is transferred to predetermined classes in these studies. These studies consist of the classification of the EEG signal. However, it has been observed that in the recent years, with the developments in the field of artificial learning, the classification has gone beyond, It can be seen that the visual viewed from the EEG signal can be produced. When the limited number of studies using classical generative adversarial networks (GAN) and autu encoder (AE) approaches are examined, it is seen that visuals from EEG signals can be produced roughly. The original aspect of this study is that it includes mathematical approaches to increase the visual production capability. Classical GAN architectures use random vector input to provide a variety of images produced. With this approach, it is observed that the visuals produced from the EEG signal are of low quality. In the proposed method, the input is considered as two parts (coded EEG and randomness). Variable auto encoder (VAE) and fourier transform (FT) are used to encode the EEG, while two different approaches are proposed for randomness. The use of this original GAN has enabled higher quality visuals to be produced than EEG signals. In order to understand this quality numerically, pre-trained convolutional neural networks (CNN) was used. As a result of experimental studies, While the performance level of the visuals produced from EEG signals with classical GAN is around 93%, it is seen that this level rises to 95% -100% in the proposed approach.

Keywords: