RF Imaging Reconstruction Project

This project implements advanced RF imaging reconstruction techniques, combining classical and neural network-based approaches. The system simulates RF measurements from 2D targets and reconstructs them using optimization and deep learning methods.

Problem Setup

• Target: 2D reflectivity distribution on the z=z₀ plane (parallel to xy plane)
• Terminals: Each terminal has a Uniform Rectangular Array (URA) transmitter and receiver, with configurable spacing, rotation, and offsets.
• Operation Modes:
  • Monostatic: Each Rx receives from its own Tx
  • Multistatic: Each Rx can receive from all Tx terminals
• Measurement Model: y = Ax, where x is the vectorized target reflectivity

Key Concepts

• Resolution Distinction:
  • Measurement Grid: High resolution (e.g., 128×128) for accurate forward modeling
  • Reconstruction Grid: Coarser resolution (e.g., 32×32 or 64×64) for computational efficiency
  • Target Bounds: 1m × 1m physical area (configurable)

Reconstruction Approaches

• Classical Methods:
  • Back Projection (BP)
  • LASSO with wavelet regularization
• VAE-Based Methods:
  • Method 1: x̂ = argmin_x ‖y - Ax‖₂² + μ‖x - ψ(x)‖₂²
  • Method 2: x̂ = G_θ(argmin_z ‖y - AG_θ(z)‖₂² + λR(G_θ(z)))
• Advanced Solvers:
  • OAMP (Orthogonal Approximate Message Passing)
  • Gradient-based methods (Adam, SGD)

Data

• Input Images: Black and white 2D images (e.g., from data_x/), used as targets for simulation and reconstruction.
• Resolutions:
  • Measurement: High (e.g., 128x128)
  • Reconstruction/Training: Coarse (e.g., 64x64)