Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Physics-Informed Wireless Imaging with Implicit Neural Representation in RIS-Aided ISAC System

Yixuan Huang, Jie Yang, Chao-Kai Wen, Xiao Li, Shi Jin

TL;DR

This work addresses wireless imaging within RIS-aided ISAC systems by introducing implicit neural representations (INR) trained with physics-informed losses. By modeling the ROI as a continuous function encoded in a six-layer MLP with Fourier positional encoding and sine activations, the method renders high-resolution images from CSI measurements without ground-truth targets or large datasets. The forward model links ROI scattering to measurements, allowing end-to-end neural optimization that avoids explicit multipath extraction and can achieve super-resolution with relatively few RIS configurations. Numerical results show near-perfect structural similarity to ground-truth targets and demonstrate advantages over Fourier/CS baselines, including a focal-distance-like behavior and robustness to measurement overhead, highlighting practical benefits for RIS-enabled ISAC imaging.

Abstract

Wireless imaging is emerging as a key capability in next-generation integrated sensing and communication (ISAC) systems, supporting diverse context-aware applications. However, conventional imaging approaches, whether based on physical models or data-driven learning, face challenges such as accurate multipath separation and representative dataset acquisition. To address these issues, this study explores the use of implicit neural representation (INR), a paradigm that has achieved notable advancements in computer vision, for wireless imaging in reconfigurable intelligent surface-aided ISAC systems. The neural network of INR is specifically designed with positional encoding and sine activation functions. Leveraging physics-informed loss functions, INR is optimized through deep learning to represent continuous target shapes and scattering profiles, enabling resolution-agnostic imaging with strong generalization capability. Extensive simulations demonstrate that the proposed INR-based method achieves significant improvements over state-of-the-art techniques and further reveals the focal length characteristics of the imaging system.

Physics-Informed Wireless Imaging with Implicit Neural Representation in RIS-Aided ISAC System

TL;DR

This work addresses wireless imaging within RIS-aided ISAC systems by introducing implicit neural representations (INR) trained with physics-informed losses. By modeling the ROI as a continuous function encoded in a six-layer MLP with Fourier positional encoding and sine activations, the method renders high-resolution images from CSI measurements without ground-truth targets or large datasets. The forward model links ROI scattering to measurements, allowing end-to-end neural optimization that avoids explicit multipath extraction and can achieve super-resolution with relatively few RIS configurations. Numerical results show near-perfect structural similarity to ground-truth targets and demonstrate advantages over Fourier/CS baselines, including a focal-distance-like behavior and robustness to measurement overhead, highlighting practical benefits for RIS-enabled ISAC imaging.

Abstract

Wireless imaging is emerging as a key capability in next-generation integrated sensing and communication (ISAC) systems, supporting diverse context-aware applications. However, conventional imaging approaches, whether based on physical models or data-driven learning, face challenges such as accurate multipath separation and representative dataset acquisition. To address these issues, this study explores the use of implicit neural representation (INR), a paradigm that has achieved notable advancements in computer vision, for wireless imaging in reconfigurable intelligent surface-aided ISAC systems. The neural network of INR is specifically designed with positional encoding and sine activation functions. Leveraging physics-informed loss functions, INR is optimized through deep learning to represent continuous target shapes and scattering profiles, enabling resolution-agnostic imaging with strong generalization capability. Extensive simulations demonstrate that the proposed INR-based method achieves significant improvements over state-of-the-art techniques and further reveals the focal length characteristics of the imaging system.
Paper Structure (28 sections, 19 equations, 6 figures, 2 tables)

This paper contains 28 sections, 19 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. System Model
  3. System Configuration
  4. Signal and Channel Models
  5. INR-Based Wireless Imaging
  6. CSI Measurements and Forward Model
  7. CSI Measurements
  8. Forward Model
  9. Network Structure
  10. Positional Encoding
  11. Activation Function
  12. NN Training and Image Generation
  13. Step 1---NN Inference
  14. Step 2---CSI Prediction
  15. Step 3---Loss Calculation
  16. ...and 13 more sections

Figures (6)

  • Figure 1: The proposed RIS-aided ISAC system for wireless imaging.
  • Figure 1: Imaging results with different system and training settings (unit for time: second).
  • Figure 2: Illustration of the INR network structure and the relationship between the image grid and the ROI image.
  • Figure 3: Example images of the synthesized dataset.
  • Figure 4: Train and test PSNR during the training process for "no RIS" and "ReLU" scenarios.
  • ...and 1 more figures