An EEG Signal Denoising Network Based on Multi-Scale Cavity Convolution

Electroencephalography (EEG) is widely used in the fields of emotion recognition, epilepsy detection and critical patient monitoring because it can provide rich physiological, psychological and pathological information. However, EEG is easily contaminated by electrooculogram (EOG) during the acquisition process, which affects the quality, reliability and interpretability of EEG, and so denoising of EEG signals is an indispensable step to analyze EEG. Currently, deep learning algorithms have shown better noise reduction performance than traditional algorithms. However, the existing deep learning algorithms have limited sensory fields, are difficult to model long-range temporal dependence, and have the problem of low computational efficiency. Therefore, this paper proposes a multi-scale Dilated Convolutional EEG Denoising Network (MS-DTNet). MS-DTNet adopts multi-scale Dilated convolution to expand the sensory field, and uses the Transformer module to complement the global features for better noise suppression. In order to evaluate the ability of MS-DTNet to suppress EOG noise, this experiment is compared with other algorithms on a public dataset. The experiments show that compared with CT-Net, MS-DTNet’s RRMSE is reduced, while CC is improved, and SNR is increased, and the above results indicate the superiority of MS-DTNet in reducing EOG noise.