CNN based Human Intention-Behavior Prediction using Common Average Reference and STFT of EEG Signals

A newly emerging research field is application of brain-computer interfaces (BCIs), for human intention-behavior prediction. In order to improve brain wave intention recognition, recent studies have included pretrained convolution neural networks (CNN). In this research, the pretrained ResNet50 model is fine-tuned using a transfer learning strategy. We first enhance the EEG signals using the common average reference (CAR) approach. Further, we compute short time Fourier transform (STFT) and obtain the spectrogram. Moreover, we combine mu and beta band images extracted from spectrogram. The combined mu and beta band images are given as the input to the pretrained ResNet50 model. The following work, we divided BCI competition-IV dataset-2b data into three ratios of training: validation: test in 6:2:2, 7:1:2, and 8:1:1. Then we performed classification using pretrained ResNet50 which resulting in classification accuracies of 91.22%, 94.42%, and 94.14%. The efficiency of the suggested model is compared with the existing models.