K-Means Based Blind Noise Variance Estimation
Esteban Selva, Apostolos A. Kountouris, Yves Louët · 2021
When an unknown signal propagates through an AWGN channel of unknown variance, estimating the noise variance can be difficult. We propose a novel method to perform blind noise variance estimation, based on the separation of noise-only values and signal-plus-noise values in the frequency representation of the received signal. This separation is conducted using the K-means algorithm. Our linear-complexity method is efficient and accurate, requires a limited amount of samples and is robust to SNRs as low as -7 dB. It relies on two weak hypotheses of compacity and sparsity on the signal of interest.