Classifying the Presence of Aircraft Images in RSVP Tasks Using EEG: A Deep Learning Approach for Addressing Class Imbalance in RSVP Tasks

In rapid serial visual presentation (RSVP) tasks, the imbalance in sample numbers can severely limit the classification performance of RSVP electroencephalography (EEG). Existing solutions typically employ rebalancing strategies (such as reweighting and resampling) to reduce the effects of class imbalance. While these strategies enhance the classifier’s learning capability (by altering the distribution of training data to focus more on samples from minority classes) and reduce the model’s bias towards majority classes, they inadvertently harm the representational power of learned deep features. To confront this challenge, this study suggests an approach that separates the feature extraction and classification processes of the model during training. This approach aims to better capture features from imbalanced EEG data while accurately classifying them. In this method, the feature extraction process is responsible for learning features from all specimens, whilst the categorizer establishes better boundaries between target and non-target classes. To assess the categorization efficacy of the proposed model in this study, experiments were conducted by comparing it with the Hierarchical Discriminant Component Analysis (HDCA) and EEGNet models on self-collected datasets. The experimental results indicate that the model demonstrates certain advantages in EEG classification for RSVP tasks. A general solution is provided to address the issue of sample imbalance.