Unsupervised Denoising via Wiener Entropy Masking in the STFT Domain

A completely blind, adaptive, and unsupervised interference nulling and denoising technique is presented. This technique works by discriminating between sub-bands belonging to narrow band pulses/pulse-trains, narrow band constant amplitude interference, and Additive White Gaussian Noise (AWGN), in the Short Time Fourier Transform (STFT) domain. Unlike traditional methods that focus solely on energy measures in the time, frequency, or joint time-frequency domains, this method utilizes the Wiener Entropy (WE) metric to isolate interference and AWGN sub-bands from those belonging to narrow band pulsed signals of interest. The Modified Wiener Entropy (MWE) can then also be used to null out sub-bands belonging to interferences in a completely threshold-less and unsupervised manner. The resulting modified STFT can then be inverse transformed back into the time-domain via the LSE method, thereby yielding a denoised instance of the original signal. As this method masks the complex-valued STFT, phase preservation of the original signal is guaranteed.