Binlets: Data fusion-aware denoising enables accurate and unbiased quantification of multichannel signals

Abstract

As monitoring multiple signals becomes more cost-effective, combining them through a data fusion-aware denoising method can produce a more robust estimation of the underlying process. Here, we present a method based on the Haar wavelet transform that trades off resolution against accuracy based on statistical significance. By taking advantage of correlations between channels, it offers a superior performance compared to denoising each channel separately. It outperforms standard wavelet methods when the magnitude of interest in the data-fusion process involves a non-linear transformation or reduction of a multichannel signal. We demonstrate its efficacy by benchmarking our method against standard wavelet thresholding for synthetic single and multichannel time series, and a multichannel two-dimensional image. The method has a simple interpretation as an adaptive binning of the signal, and neither requires training data nor specialized hardware to run fast. In addition, a reference Python implementation is available on GitHub and PyPI, making it simple to integrate into any analysis pipeline.