EEG emotion recognition based on 2DSinc convolution and attention mechanism

In addressing the issues of low accuracy, Feature extraction capabilities are limited, and insufficient interpretability associated with conventional convolutional neural networks (CNN) in EEG emotion recognition tasks, this study introduces a novel EEG emotion recognition method inspired by the SincNet network architecture. This method also considers the limitation of SincNet, which is primarily applicable to single-channel audio signals. After preprocessing and segmenting the EEG signals, a parametric 2D Sinc layer is utilized to capture the essential features of the input signal. Local features are then further extracted using depthwise convolution. To improve the model’s ability to emphasize crucial information, an attention mechanism is incorporated to extract global features along the temporal dimension. Lastly, a fully connected layer is employed to complete the four-class classification task. Experimental results reveal that the proposed model achieves a classification accuracy of $95.05 \%$ in subjectdependent experiments on the DEAP dataset, and $86.72 \%$ in crosssubject experiments. These results significantly outperform conventional CNN methods, The effectiveness of the proposed method for EEG emotion recognition tasks is confirmed.