μD-Diff: A Universal Blind Denoising Method for Radar Spectrograms Using Diffusion Model

Radar-based human motion sensing has significantly advanced. However, environmental noise and interference bring about new challenges, particularly in the blind denoising scenario where it is impossible to predict the noise level beforehand. Current radar blind denoising techniques mostly rely on generative adversarial networks (GANs), which may have issues with training instability and thus be incapable of producing reliable denoised spectrograms. This article proposes a universal blind denoising model named “ $\mu $ D-Diff.” Based on the diffusion probabilistic model (DDPM), we develop a radar noise guidance mechanism to provide more accurate distribution fitting ability and better retain semantic information. In addition, a step control module (SCM) is introduced in the model to adaptively choose an optimal sampling step for the diffusion model and seek a balance between under-denoising and over-denoising. Experimental results demonstrate that our proposed method outperforms several state-of-the-art radar spectrogram denoising methods and shows superior generalization ability over various testing scenarios.