EEG-WNet: A wavelet-guided dual-branch network for ocular artifact suppression in electroencephalogram

Electroencephalogram (EEG) artifact removal is essential for reliable neural signal analysis, yet many existing methods do not balance reconstruction quality and efficiency well enough for real-time use. EEG-WNet is a lightweight dual-branch framework that combines time-domain modeling with wavelet-guided prior learning for EEG artifact removal. Cross-attention fusion, residual-guided refinement, and a lightweight multi-scale channel-spatial attention module are used to improve reconstruction quality without relying on aggressive model scaling. Experiments on semi-synthetic datasets focus on electrooculographic (EOG) contamination as the primary task and include additional evaluations on electromyographic (EMG) and synthetic noise contamination. On EEGdenoiseNet, EEG-WNet achieves an RRMSE t of 0.212, an SNR of 15.42 dB, and a CC of 0.967 for EOG removal while remaining competitive across other artifact types. Qualitative transfer results on DEAP further indicate reasonable robustness on real EEG recordings. The model requires 8.48 M parameters, 48.90 M FLOPs, and 1.34 ms CPU inference time, supporting its use as a lightweight front-end module for EEG preprocessing.