Artifact Mitigation for High-Resolution Near-Field SAR Images by Means of Conditional Generative Adversarial Networks

Abstract

This work presents an approach to enhance the quality of high-resolution images obtained by means of systems relying on synthetic aperture radar (SAR). For this purpose, a deep learning method called conditional generative adversarial networks (cGAN) is applied to the imager outcome when it is prone to suffer artifacts. This is specially the case of novel systems pushing the limits of SAR (e.g., irregular sampling, multilayered media, etc.) resulting in very chaotic clutter and image artifacts that cannot be easily removed with conventional approaches. The cGAN can be trained to detect high-level characteristic features in the image (e.g., parts of a scissor blade) so another output based on these detected features can be tailored. In other words, it can translate features contaminated by artifacts into clean features, effectively improving the quality of SAR images. Unlike other deep learning approaches, the training of the involved neural networks tends to be stable thanks to the structure based on two competing subsystems. The proposed approach is illustrated using simulated and measurement data in the context of two advanced near-field SAR systems considering: i) cylindrical multilayered media, and ii) freehand acquisitions. Results show that cGANs clearly outperform conventional approaches removing most of the artifacts, enabling to produce a clean output image.

Single Radar Image Motion Deblurring Through PSF Estimation and Convolution Regularizer

“…[46] focused on a 2D image denoising model using complexvalued CNNs and residual networks, yielding a deep network with a higher parameter to solve the denoising problem. In [47], a complex-valued conditional GAN (cGAN) model was developed to enhance radar image quality. Results showcased significant advancements over conventional methodologies by effectively eliminating artifacts present in the input images.…”

DI-Gesture: Domain-Independent and Real-Time Gesture Recognition with Millimeter-Wave Signals

“…While these methods have explicit theoretical foundations, they would face challenges in near-field 3D mmWave imaging because the optimization difficulty and computational cost for a large-scale array (usually tens of thousands of elements) are unacceptable. To address these issues, recent studies [37,38] shift their attention to learning-based methods, demonstrating huge success in image processing, for handheld SAR imaging. However, since these methods are based on neural networks designed for processing real-valued RGB images, they only take the intensity of complex-valued SAR images as the input while discarding the phase information.…”

“…For instance, some researchers [37] proposed to employ an efficient convolutional neural network for freehand SAR image superresolution. Others [38] leveraged conditional generative adversarial networks to mitigate the image artifact caused by positioning errors. While these methods reported good results on corresponding datasets, they failed to incorporate the characteristics of mmWave images (e.g.…”

A high-resolution handheld millimeter-wave imaging system with phase error estimation and compensation

“…Nonetheless, these approaches still depend on the premise of accurate device tracking, which is a non-trivial problem. In addition to signal processing-based methods, recent studies 29 , 33 , 34 have directly input the amplitude image into deep neural networks to alleviate distortions caused by motion errors. However, they ignore the signal phase critical for phase error compensation and simplify it as an image super-resolution problem.…”