P300 Evoked Related Potential Detection Based on Integration of Modified HOG and Convolutional Neural Networks

This paper proposes a novel method based on the combination of modified histogram of oriented gradients (MHOG) and convolutional neural network (CNN) for P300 evoked related potential (ERP) detection. In this method, HOG is modified for 2-D EEG signal and is fed to CNN as input. Due to the inputs as MHOG, the training features are chosen among a variety of gradients, and the best features are selected for training the CNN structure. This approach (MHOG-CNN) ends in a simpler structure and therefore, the whole performance is acceptable with faster rates and high accuracy. The performance of MHOG-CNN is examined by applying BCI Competition III Dataset. This dataset represents a complete record for P300 ERP with BCI2000 using a paradigm, and it has several noises including power and muscle-based noises. The objective is to predict the correct character in each of the provided character selection epochs. In comparison to other methods, simulation results indicate remarkable abilities of MHOG-CNN structure for P300 ERP detection. Our new method yields the classification rates over 97.12% with comparable execution time for the mentioned dataset.