[Paper Review] Mitigation of Through-Wall Distortions of Frontal Radar Images using Denoising Autoencoders

Shelly Vishwakarma, Shobha Sundar Ram|arXiv (Cornell University)|Mar 22, 2019

Geophysical Methods and ApplicationsEngineering42 references29 citations

TL;DR

This paper proposes a denoising autoencoder (DAE) based on the alternating direction method of multipliers (ADMM) to mitigate through-wall radar image distortions without prior knowledge of wall parameters or propagation models. Trained on diverse through-wall and free-space radar image pairs, the DAE effectively restores frontal radar images of static and dynamic humans, achieving high structural similarity and low mean square error compared to clean reference images.

ABSTRACT

Radar images of humans and other concealed objects are considerably distorted by attenuation, refraction and multipath clutter in indoor through-wall environments. While several methods have been proposed for removing target independent static and dynamic clutter, there still remain considerable challenges in mitigating target dependent clutter especially when the knowledge of the exact propagation characteristics or analytical framework is unavailable. In this work we focus on mitigating wall effects using a machine learning based solution -- denoising autoencoders -- that does not require prior information of the wall parameters or room geometry. Instead, the method relies on the availability of a large volume of training radar images gathered in through-wall conditions and the corresponding clean images captured in line-of-sight conditions. During the training phase, the autoencoder learns how to denoise the corrupted through-wall images in order to resemble the free space images. We have validated the performance of the proposed solution for both static and dynamic human subjects. The frontal radar images of static targets are obtained by processing wideband planar array measurement data with two-dimensional array and range processing. The frontal radar images of dynamic targets are simulated using narrowband planar array data processed with two-dimensional array and Doppler processing. In both simulation and measurement processes, we incorporate considerable diversity in the target and propagation conditions. Our experimental results, from both simulation and measurement data, show that the denoised images are considerably more similar to the free-space images when compared to the original through-wall images.

Motivation & Objective

• To address through-wall radar image distortions caused by attenuation, refraction, and multipath clutter.
• To develop a machine learning solution that does not require prior knowledge of wall characteristics or analytical propagation models.
• To enable robust image restoration using only paired through-wall and free-space training data.
• To ensure fast convergence and real-time applicability through ADMM-based optimization.

Proposed method

• Trained a denoising autoencoder using paired through-wall and line-of-sight (free-space) radar images as input and output.
• Employed an ADMM-based optimization framework to accelerate training and guarantee convergence.
• Used wideband planar array data for static targets and narrowband data with Doppler processing for dynamic targets.
• Incorporated stochastic variations in FDTD simulations to model diverse wall parameters (dielectric constant, conductivity, reinforcements, air gaps).
• Applied two-dimensional array and range/Doppler processing to generate frontal radar images from raw radar data.
• Optimized hyperparameters including hidden layer size and activation functions (linear, tanh, sigmoid) via empirical evaluation.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively reduce through-wall distortions in frontal radar images without prior knowledge of wall properties?
• RQ2 How well does a denoising autoencoder generalize across diverse wall types and target conditions?
• RQ3 Does ADMM-based training enable faster convergence and real-time deployment compared to standard backpropagation?
• RQ4 How does model performance vary with different activation functions and hidden layer sizes?
• RQ5 To what extent can the model tolerate labeling mismatches between free-space and through-wall image pairs?

Key findings

• The denoised images showed significantly lower normalized mean square error (NMSE) and higher structural similarity (SSIM) compared to original through-wall images.
• For simulation data, the linear activation function outperformed non-linear functions (tanh, sigmoid), likely due to the narrowband nature of the data.
• With 500 hidden nodes, SSIM converged in simulations, indicating sufficient capacity for effective denoising.
• For real measurement data, SSIM approached 1.0 with 1500 hidden nodes, with the sigmoid function yielding the best performance.
• Test inference time remained low (under 500MB memory, feasible on Raspberry Pi 3+), enabling real-time deployment.
• The model demonstrated robustness to diverse wall conditions and target variations without requiring wall-specific tuning.