Learning from Natural Noise to Denoise Micro-Doppler Spectrogram

Chong Tang, Wenda Li, Shelly Vishwakarma, K. Woodbridge, Simon Julier, Kevin ChettyarXiv 2021 · DOI · arXiv

Abstract

Micro-Doppler analysis has become increasingly popular in recent years owing to the ability of the technique to enhance classification strategies. Applications include recognising everyday human activities, distinguishing drones from birds, and identifying different types of vehicles. However, noisy time-frequency spectrograms can significantly affect the performance of the classifier and must be tackled using appropriate denoising algorithms. In recent years, deep learning algorithms have spawned many deep neural network-based denoising algorithms. For these methods, noise modelling is the most important part and is used to assist in training. In this paper, we decompose the problem and propose a novel denoising scheme: first, a Generative Adversarial Network (GAN) is used to learn the noise distribution and correlation from the real-world environment; then, a simulator is used to generate clean Micro-Doppler spectrograms; finally, the generated noise and clean simulation data are combined as the training data to train a Convolutional Neural Network (CNN) denoiser. In experiments, we qualitatively and quantitatively analyzed this procedure on both simulation and measurement data. Besides, the idea of learning from natural noise can be applied well to other existing frameworks and demonstrate greater performance than other noise models.

Introduction

Micro-Doppler analysis is a powerful tool for wireless sensing, enabling systems to identify human activities or distinguish objects by analyzing the minute frequency shifts caused by moving components. However, real-world signals, particularly from passive WiFi radar, are often obscured by complex, structured noise.

Standard deep learning approaches to denoising require pairs of noisy and perfectly clean data for training. While clean signals can be simulated effectively, simulating realistic noise remains a challenge. Most researchers rely on Additive White Gaussian Noise (AWGN), but real-world noise is spatially and temporally correlated. If a denoiser is trained only on AWGN, it often fails to generalize to the patterned noise of real-world environments. This paper introduces a novel framework that replaces simplistic artificial noise models with a Generative Adversarial Network (GAN) trained on "natural" noise extracted from real-world recordings.

Methodology

The proposed pipeline follows a "decompose and recombine" strategy consisting of three primary stages:

  1. Noise Modeling (The GAN): Instead of mathematical formulas, a Wasserstein GAN with Gradient Penalty (WGAN-GP) is employed. The model is fed "noise-only" patches extracted from idle periods in real radar recordings. This allows the GAN to learn the specific spatial and temporal correlations of the environment.
  2. Signal Simulation: The SimHumalator tool generates clean Micro-Doppler Spectrograms (MDS) based on human motion capture (Mocap) data.
  3. The Denoiser (DnCNN): A training set is created by superimposing the GAN-generated realistic noise onto the simulated clean signals. A Denoising Convolutional Neural Network (DnCNN) is then trained on these pairs. Crucially, the DnCNN uses residual learning to predict the noise patch itself, which is then subtracted from the input to reveal the clean signal.
The proposed three-stage denoising framework pipeline.

Datasets

  • Real-World Data: Collected using a Passive WiFi Radar (PWR) system in a living room setting, featuring 360 measurements of activities like walking and falling.
  • Simulated Data: 360 datasets generated via SimHumalator to match the real-world activities.
  • Noise Patches: Extracted from "idle periods" between activities in real radar recordings to provide the GAN with authentic environmental noise distributions.

Key Insights and Results

A central finding of this research is that real radar noise has "texture"—often appearing as narrow rectangular patterns rather than random pixels. By training the GAN to capture this texture, the resulting denoiser becomes significantly more effective at identifying and removing environmental interference without blurring the underlying signal.

Quantitative and Qualitative Performance

In simulations, the Measurement Noise Model (MNM) approach consistently outperformed AWGN-based models, achieving higher Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) across noise levels ranging from -10dB to 25dB.

Qualitative comparison of denoising results: (a) Ground truth, (b) Noisy input, (c) AWGN-based denoising, (d) Proposed MNM-based denoising.

Furthermore, the denoiser's impact on downstream classification was tested using a VGG16 architecture. At a 10dB noise level, the MNM-trained denoiser improved activity classification accuracy to 97%, a 2% improvement over the baseline and a significant jump over models trained on simplistic noise distributions.

Conclusion

By learning from natural noise rather than relying on mathematical approximations, this framework provides a more robust solution for Micro-Doppler signal processing. While the current noise model is environment-specific, the methodology demonstrates that focusing on noise distribution learning bypasses the need for impossible-to-obtain real-world "clean" labels. Future work will explore the adaptability of these models across varying environmental conditions and dynamic scenarios.

Summary updated

Abstract

Micro-Doppler analysis has become increasingly popular in recent years owning to the ability of the technique to enhance classification strategies. Applications include recognising everyday human activities, distinguishing drone from birds, and identifying different types of vehicles. However, noisy time-frequency spectrograms can significantly affect the performance of the classifier and must be tackled using appropriate denoising algorithms. In recent years, deep learning algorithms have spawned many deep neural network-based denoising algorithms. For these methods, noise modelling is the most important part and is used to assist in training. In this paper, we decompose the problem and propose a novel denoising scheme: first, a Generative Adversarial Network (GAN) is used to learn the noise distribution and correlation from the real-world environment; then, a simulator is used to generate clean Micro-Doppler spectrograms; finally, the generated noise and clean simulation data are combined as the training data to train a Convolutional Neural Network (CNN) denoiser. In experiments, we qualitatively and quantitatively analyzed this procedure on both simulation and measurement data. Besides, the idea of learning from natural noise can be applied well to other existing frameworks and demonstrate greater performance than other noise models.
BibTeX
@misc{tang2021learning,
  author = {Chong Tang and Wenda Li and Shelly Vishwakarma and K. Woodbridge and Simon Julier and Kevin Chetty},
  title = {{Learning from Natural Noise to Denoise Micro-Doppler Spectrogram}},
  year = {2021},
  eprint = {2102.06887},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.06887},
}
Figure