| Popis: |
In this paper, guided by the prior knowledge that the overlap degrees of EEGs and EMGs vary across frequency bands, we propose a dual-stream temporal convolutional network (TCN) with attentions for EMG removal from single-channel EEG, called DSATCN. DSATCN consists of two groups of 1-D convolution-based encoder and decoder pairs and TCN-based separators. The first group extracts the reliable features of EEG recordings in a low-frequency band, which contains fewer EMG artifacts. The second group adaptively fuses these reliable features by the channel-attention network to improve the feature extraction and mask learning of the full frequency band EEG. Furthermore, regarding the diversity of latent EEG dependencies, we add the inception layers in TCNs to extract the multi-scale EEG features, employ multi-head self-attentions to gather global information in multi-levels, and use the channel-temporal wise dropout to improve the generalization performance. And we use two public datasets to evaluate the effectiveness of DSATCN, which are EEGdenoiseNet and BioSource EEG. |
