DenoFlow
A flow matching-based EEG denoising model designed for SSVEP brain-computer interface systems. It formulates the denoising problem as a conditional transport process and introduces a decodability supervision objective to guide signal reconstruction.
Timeline (last 90 days)
DenoFlow is a method for denoising steady-state visual evoked potential (SSVEP) EEG signals, adopting a Flow Matching modeling approach.
DenoFlow was published on arXiv.
DenoFlow redefines the denoising problem as a transport process rather than learning a direct mapping from corrupted signals to clean signals.
DenoFlow adopts the rectified-flow framework, where the field network regresses the velocity field along the straight-line path between corrupted and clean signals.
DenoFlow uses the observed corrupted signal as the starting point for integration rather than random noise, thereby inheriting the structural information of the original signal and reducing the risk of signal distortion.
DenoFlow's training process is simplified to a regression problem, avoiding the unstable min-max game in adversarial generative networks.
DenoFlow's field network adopts an encoder-decoder architecture, where the corrupted signal is visible to the network at every layer and path position information is injected into the bottleneck layer.
DenoFlow trains a classifier in parallel to supervise the integrated output, incorporating whether the output remains decodable as an optimization objective into training.
DenoFlow artificially constructs paired training samples by injecting real-recorded EMG and EOG physiological artifacts into clean signals at controlled signal-to-noise ratios.
DenoFlow was tested on two public SSVEP datasets in combination with five mainstream SSVEP decoders, outperforming seven baseline denoising models on both signal fidelity and downstream decoding accuracy.
1 more timeline events
All Facts (11)
DenoFlow is a method for denoising steady-state visual evoked potential (SSVEP) EEG signals, adopting a Flow Matching modeling approach.
50%UnverifiedDenoFlow was published on arXiv.
50%UnverifiedDenoFlow redefines the denoising problem as a transport process rather than learning a direct mapping from corrupted signals to clean signals.
50%UnverifiedDenoFlow adopts the rectified-flow framework, where the field network regresses the velocity field along the straight-line path between corrupted and clean signals.
50%UnverifiedDenoFlow uses the observed corrupted signal as the starting point for integration rather than random noise, thereby inheriting the structural information of the original signal and reducing the risk of signal distortion.
50%UnverifiedDenoFlow's training process is simplified to a regression problem, avoiding the unstable min-max game in adversarial generative networks.
50%UnverifiedDenoFlow's field network adopts an encoder-decoder architecture, where the corrupted signal is visible to the network at every layer and path position information is injected into the bottleneck layer.
50%UnverifiedDenoFlow trains a classifier in parallel to supervise the integrated output, incorporating whether the output remains decodable as an optimization objective into training.
50%UnverifiedDenoFlow artificially constructs paired training samples by injecting real-recorded EMG and EOG physiological artifacts into clean signals at controlled signal-to-noise ratios.
50%UnverifiedDenoFlow was tested on two public SSVEP datasets in combination with five mainstream SSVEP decoders, outperforming seven baseline denoising models on both signal fidelity and downstream decoding accuracy.
50%UnverifiedDenoFlow's code has been open-sourced on GitHub at https://github.com/wzwvv/DenoFlow.
50%