A practical low-dimensional feature vector generation method based on wavelet transform for psychophysiological signals

Abstract

: High-dimensional feature vectors entail computational cost and computational complexity. However, a successful classification can be obtained with an optimally sized feature vector consisting of distinctive features. With the widespread use of the internet and mobile devices, the need for systems with low computational costs is increasing day by day. In this study, starting from the idea that each motor imagery is represented as a subject-specific pattern in the brain, we propose a new and practical method that can generate a low-dimensional feature vector based on wavelet transform. The feature vector is obtained from the correlation between each trial and each class average. To investigate the effect of possible temporal shifts in the trial signals, the proposed method is analyzed with signal segments with different starting points and lengths. The effect of these signal segments on classification is shown. The proposed feature extraction approach is tested on two different datasets and the classification results are presented in comparison with previous studies. With the method proposed in this study, much lower-dimensional feature vectors are obtained compared to previous studies and very satisfactory results are obtained. It is observed that EEG signals related to motor imagery in the brain have a subject-specific pattern, and this pattern is successfully classified with a feature vector consisting of only 1 feature per class.