Frequency Information Enhanced Deep EEG Denoising Network for Ocular Artifact Removal

Abstract

Electroencephalography (EEG) signals are easily contaminated by various artifacts, making noise removal an essential step in EEG analysis. In recent years, deep-learning-based methods have provided a promising avenue for EEG denoising, with superior performance on several benchmarks. However, most existing deep models only rely on the temporal features of the signal and ignore the intrinsic characteristics in other domains, possibly limiting their ability to suppress artifacts. Therefore, we propose a cross-domain framework by integrating knowledge of both time and frequency domains into a deep-learning model for ocular artifact removal from EEG recordings. The proposed framework can be flexibly implemented on various deep denoising networks to improve their denoising performance. Extensive experiments demonstrate that the cross-domain design is capable of better eliminating the ocular artifact when implemented over several representative denoising networks. For example, RRMSE is reduced by 4%–26% for semisimulated data. These results verify the effectiveness of the proposed framework, which also indicates the importance of combining domain knowledge into the deep-learning models for EEG denoising.