⚡ electrical engineering

Generative Adversarial Reconstruction with Adaptive Thresholding for Obstructed Targets in Computational Microwave Imaging

This paper proposes an end-to-end generative adversarial framework integrating a conditional GAN with a learnable soft-threshold module to effectively reconstruct targets obstructed by undesired objects in computational microwave imaging, demonstrating robust performance across diverse datasets, experimental measurements, and varying signal-to-noise ratios.

Original authors: Jiaming Zhang, Maria Garcia-Fernandez, Guillermo Alvarez-Narciandi, Jie Zhang, Muhammad Ali Babar Abbasi, Okan Yurduseven

Published 2026-08-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiaming Zhang, Maria Garcia-Fernandez, Guillermo Alvarez-Narciandi, Jie Zhang, Muhammad Ali Babar Abbasi, Okan Yurduseven

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to take a clear photo of a secret treasure chest, but someone has piled a messy heap of junk right on top of it. In the world of science, this is a bit like "microwave imaging." Instead of using light to see things, scientists use invisible microwave waves to "see" through walls, debris, or even clothes. It's a superpower used for everything from finding survivors after an earthquake to checking for hidden objects in security scans. Usually, these systems work by bouncing waves off an object and using a computer to draw a picture of what's underneath. But here's the catch: if there's a pile of junk (or "obstructions") in the way, the computer gets confused. It tries to draw the junk and the treasure together, resulting in a blurry, messy picture that's hard to read. For a long time, the only way to fix this was to physically move the junk away, which is slow, difficult, and sometimes impossible.

Now, picture a team of scientists who decided to teach a computer to be a "mental magician." They didn't want to move the junk; they wanted the computer to learn how to ignore it and only show the treasure. This is exactly what the researchers in this paper have done. They created a new, smart system called cGAN-STM that acts like a digital filter. Think of it as a pair of glasses that automatically blurs out the background noise and sharpens the important parts, but it does this by "learning" what the junk looks like and what the target looks like, rather than just following a manual rulebook.

Here is how their magic trick works. They built a system with two main parts that play a game against each other, much like a forger trying to make a perfect fake painting and an art critic trying to spot the fake. The "forger" (called a generator) looks at the messy microwave signals bouncing off the blocked target and tries to guess what the clean target looks like underneath. The "critic" (called a discriminator) checks the guess to see if it looks real. They keep playing this game over and over, getting better at every round. But there's a special twist: they added a "soft-threshold module" (STM). You can think of this module as a very smart bouncer at a club. When the signal comes in, the bouncer checks every piece of information. If a piece of data looks weak or like it belongs to the "junk" (the obstruction), the bouncer gently kicks it out. If it looks strong and important (like the target), the bouncer lets it pass through. The best part? The bouncer doesn't need a human to tell it what to kick out; it learns the perfect rules on its own while playing the game.

The researchers tested this idea in two ways. First, they used a massive computer simulation with thousands of digital examples. They created a scenario where handwritten numbers (like the digits 0 through 9) were the "treasure," and handwritten letters were the "junk" piled on top. They fed the messy microwave signals of these blocked numbers into their new system. The results were impressive: the system successfully stripped away the letters and reconstructed the numbers with high clarity. In these simulations, the new method produced images that were 87.6% structurally similar to the perfect, unblocked images (a score known as SSIM), and the error rate was very low at 0.066. To put that in perspective, when they tried to use older, standard methods on the same blocked targets, the images were so blurry that the numbers were barely recognizable.

But the scientists didn't stop at just computer games. They took their system to a real-world lab and tested it with actual microwave equipment. They placed real objects in front of each other and measured the signals. Even with real-world noise and imperfections, the system managed to reconstruct the hidden targets. When they added artificial "static" (noise) to the signals to simulate a messy environment, the system still worked well. Even when the signal was quite noisy (at a level of 15 dB), the system could still produce a recognizable outline of the target, with an error rate of 0.168 and a similarity score of 0.700. This suggests that the system is robust enough to handle real-life challenges, not just perfect lab conditions.

The paper also explored how well this "mental magician" could handle different situations. They tested it with different sizes of "junk" (obstructions) and found that as the junk got bigger and covered more of the target, the image quality did drop, but the system still tried its best to recover what it could. They even swapped the roles, using letters as the treasure and numbers as the junk, and the system adapted perfectly, proving it wasn't just memorizing one specific trick. Perhaps most excitingly, they showed that the system could be quickly "calibrated" to recognize completely new shapes, like triangles or squares, using just a tiny amount of new data, suggesting it could be adapted for many different types of objects in the future.

In short, this paper suggests that we don't always need to physically clear the path to see what's behind it. By teaching a computer to intelligently filter out the noise and focus on the signal, we can reconstruct clear images of targets even when they are heavily blocked. While the system isn't perfect—it struggles a bit when the obstruction is huge or the noise is extreme—it represents a significant step forward. It offers a way to make microwave imaging faster and more efficient, potentially changing how we do security checks or search for objects in cluttered environments, all by letting the computer learn to ignore the mess.

Technical Summary: Generative Adversarial Reconstruction with Adaptive Thresholding for Obstructed Targets in Computational Microwave Imaging

Problem Statement
Computational Microwave Imaging (CMI) offers a hardware-efficient alternative to traditional Synthetic Aperture Radar (SAR) by utilizing spatially incoherent radiation patterns to compress data acquisition. However, a significant limitation in practical CMI applications is the presence of obstructing objects (clutter) that occlude the targets of interest. In such scenarios, the back-scattered measurements contain mixed information from both the target and the obstruction. Conventional reconstruction methods, such as least-squares inversion, fail to separate these components, resulting in poor-quality images where the target cannot be accurately identified. Furthermore, physically removing obstructions or manually filtering data is often time-consuming, inefficient, or impractical. While deep learning has been applied to microwave imaging, standard feed-forward networks often produce averaged solutions with poor perceptual quality, and existing Generative Adversarial Network (GAN) approaches typically assume unobstructed imaging scenarios.

Methodology
The authors propose an end-to-end framework termed cGAN-STM (conditional Generative Adversarial Network with Soft-Thresholding Module). This framework is designed to reconstruct targets of interest directly from back-scattered measurements of obstructed scenes, effectively treating obstruction-related information as task-specific noise.

Architecture:
- Generator: Based on a U-Net architecture, the generator transforms complex-valued back-scattered measurements (magnitude and phase) into image reconstructions. It employs an encoder-decoder structure with skip connections. Crucially, the decoder incorporates Soft-Thresholding Modules (STMs).
- Soft-Thresholding Module (STM): Unlike traditional fixed-thresholding, the STM is a learnable sub-network. It utilizes global average pooling, fully connected layers, and a sigmoid activation to automatically generate optimal threshold values ( $\tau$ ) for each input. These thresholds are applied via a soft-thresholding function to suppress weak activations (likely associated with obstructions or noise) while preserving strong, salient target features.
- Discriminator: A convolutional network that distinguishes between generated reconstructions and ground-truth reconstructions, enforcing visual realism through adversarial training.
Training Strategy:
- The model is trained using a composite loss function comprising an adversarial loss (Binary Cross-Entropy) and an $L_1$ -norm loss to ensure pixel-wise similarity with the ground truth.
- The training dataset consists of synthetic back-scattered measurements generated using an experimentally characterized sensing matrix. The targets of interest are handwritten digits (MNIST), and the obstructions are handwritten letters (E-MNIST), randomly positioned and scaled to create diverse occlusion patterns.
- The framework is trained to map obstructed measurements directly to the clean target reconstruction space, bypassing the need for a separate obstruction removal step.

Key Contributions

Novel Framework: The paper presents a proof-of-concept implementation of a deep generative adversarial learning framework specifically for CMI scenarios involving obstructed targets.
Adaptive Filtering: The development of the learnable STM allows the network to adaptively filter out obstruction-induced features without manual threshold selection, a process that is typically time-consuming and suboptimal.
Direct Reconstruction: The approach eliminates the need for manual intervention to separate target information from obstructions, reducing computational costs compared to traditional iterative inversion methods.
Comprehensive Validation: The study includes extensive benchmarking against state-of-the-art methods (including Att-GAN, DCCGAN, ASR-T, U-Net, and V-Net), ablation studies, and validation using both synthetic data and experimental measurements from a physical CMI system.

Results

Quantitative Performance: Under ideal (noise-free) conditions, the proposed cGAN-STM achieved a Normalized Mean Square Error (NMSE) of 0.066 and a Structural Similarity Index (SSIM) of 0.876 on the testing dataset. This significantly outperformed conventional least-squares methods and other deep learning benchmarks, which yielded higher NMSE and lower SSIM values when applied to obstructed targets.
Experimental Verification: The framework was tested using an experimental CMI system operating in the X-band (8–12 GHz). The results demonstrated the feasibility of the approach in practical settings, successfully reconstructing targets obscured by letters.
Robustness and Adaptability:
- Noise: The model maintained recognizable target outlines at low Signal-to-Noise Ratios (SNR), achieving an NMSE of 0.168 and SSIM of 0.700 at 15 dB SNR.
- Obstruction Size: Performance degraded as obstruction size increased (from 60% to 90% of the target size), but the model remained effective for moderate occlusions.
- Generalization: The model demonstrated adaptability to unseen geometric shapes (disks, squares, triangles) and different reflectivity levels after lightweight calibration with a small dataset (300 samples).
- Ablation: Removing the STM (cGAN-STM-Lite) resulted in lower reconstruction quality, confirming the module's critical role in feature filtering.

Significance and Claims
The authors claim that cGAN-STM effectively addresses the challenge of reconstructing targets in the presence of obstructions, a scenario where conventional CMI methods fail. By learning to suppress non-target features and infer missing target information, the framework offers a robust solution that enhances the efficiency of CMI systems. The paper highlights that this approach could be particularly beneficial for applications requiring rapid, automated imaging without manual preprocessing, such as security screening. The study concludes that the proposed method is a significant step toward scalable and adaptable CMI systems capable of handling complex, real-world imaging conditions.

More like this