Enhancing RSVP Target Recognition Performance by Multifeatures Extracted From EEG Source Localization and Multigranular Information Neural Networks

The brain-computer interface (BCI) based on rapid serial visual presentation (RSVP) exhibits enhanced accuracy in target recognition in complex environments by effectively extracting electroencephalogram (EEG)'s features, thereby it may compensate for recognition defects of computer vision-based vision measurement systems. Given the current limited comprehension of the brain in existing BCI systems and the suboptimal decoding accuracy in the RSVP target recognition system, this study constructs a source model for an in-depth investigation into the brain's response mechanism, extracts fusion features based on deep learning (DL) to improve the system decoding accuracy, and establishes a theoretically based, high-precision target recognition model. The source model inversely calculates cortical neural activity by EEG and extracts significantly distinguishable features. A DL network based on a multigranular information network (MGIFNet) is introduced to obtain deep feature representation of cortical signals and connect the depth feature with other source features in series to form the fusion feature based on cortical signals. Experimental results show that, in the RSVP target recognition tasks, the area under the curve (AUC) and true positive rate (TPR) of the source features-based model are 23.39% and 29.44% higher than that of EEG-based EEGNet. With multigranular information (MGI) data representation, the average AUC and TRP of fusion feature classification can reach 93.68% and 91.94%. This indicates that the introduction of source localization and the MGIFNet provides theoretical support for feature extraction and classification results, which significantly improving target recognition performance.