ScholarWorks@성균관대학교

초록

Electroencephalogram (EEG) signals serve as a primary input for brain-computer interface (BCI) systems, and extensive research has been conducted on EEG-based emotion recognition. However, because EEG signals are inherently contaminated with various types of noise, the performance of emotion recognition is often degraded. Furthermore, the use of a Band Feature Extraction Neural Network (BFE-Net), a state-of-the-art (SOTA) method in this field, has limitations with respect to independent band-wise feature extraction and a simplistic band aggregation process to obtain final classification results. To address these problems, this study proposes the noise-robust band-attention BFE-Net framework, aiming to improve the conventional BFE-Net from two perspectives. First, we implement multiple-input, multiple-output (MIMO)-based preprocessing. Specifically, we utilize multichannel minima-controlled recursive averaging for precise non-stationary noise covariance estimation and generalized eigenvalue decomposition for subspace filtering to enhance the signal-to-noise ratio. Second, we propose an attention-based band aggregation mechanism. By integrating a band-wise self-attention mechanism, the model learns dynamic inter-band dependencies for more sophisticated feature fusion for classification. Experimental results on the SEED and SEED-IV datasets under a subject-independent protocol show that our model outperforms the SOTA BFE-Net by 3.27% and 3.34%, respectively. This confirms that rigorous MIMO noise reduction, combined with frequency-centric attention, significantly enhances the reliability and generalization of BCI systems.

키워드