Classification of EEG-based motor imagery tasks using 2-D features and quasi-probabilistic distribution models

Motor imagery (MI) is one of the most frequently used ways of designing brain-computer interfaces (BCIs). In this method, motor actions are performed by stimulating the brain without using the neuromuscular system. In this dissertation, new pattern recognition techniques are presented for the classification of MI tasks. For this purpose, firstly, transition points on the signals are detected and 2-D features are extracted in a specialized feature space. After that, learning models are constructed using 2-D quasi-probabilistic distribution models (QPDM) and classification is carried out with probabilistic memberships. In the following, QPDM-based sub-classifiers are designed using data domain and some classifiers that also adopt the combinations in voting methods are proposed. Moreover, 2-D features are transformed into 2-D modelling images, and in this way, EEG signals are given as input to convolutional neural networks with the assistance of transfer learning. In addition to all these studies, a new data set, abbreviated as MI-BMPI, is introduced that contains MIs of most frequently used hand movements on mobile phones, and thus brain-mobile phone interfaces are added to the problem domain. The performances of proposed methods are tested on MI-BMPI and several BCI Competition data sets.