A Neural Network-Assisted Denoiser for Sparse Signals With Low Rank Property of Transformed Hankel Matrices

Abstract

Designing a denoising framework for high-mobility environments is challenging due to the limited size of collected data and low latency requirements. In this paper, we introduce a neural network (NN)-assisted denoiser for sparse signals in the frequency domain, referred to as dssNET, based on the low-rank property of the transformed Hankel matrices constructed from sparse signals. The proposed method is based on optimizing the NN model by inputting singular values of the noisy transformed Hankel matrices and outputting the ground truth singular values. Furthermore, we additionally propose the advanced version of dssNET, referred to as selective dssNET (sdssNET), which can be operated more adaptively with the current signal-to-noise ratio (SNR). Notably, the proposed schemes show excellent denoising performance while requiring an extremely small training dataset compared to conventional schemes. Finally, we provide an application of joint-range-and-velocity estimation in automotive radar systems to validate the benefit of our proposed method in practical scenarios.