Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PISE: Physics-Anchored Semantically-Enhanced Deep Computational Ghost Imaging for Robust Low-Bandwidth Machine Perception

Tong Wu

TL;DR

This work tackles robust machine perception from severely undersampled computational ghost imaging measurements, where the forward model is $\mathbf{y}=A\mathbf{x}+\boldsymbol{\eta}$ with $A\in\mathbb{R}^{M\times N}$ and sampling rate $\gamma = M/N$. It proposes PISE, which combines physics-anchored initialization via $\mathbf{x}_{init}=\mathcal{R}(A^T\mathbf{y})$ and semantically guided reconstruction through a frozen VGG-based perceptual loss $\mathcal{L}_{perc}$ balanced with an MSE term, along with a gradient dynamics metric $G(t)=\left\|\nabla_{\theta}\mathcal{L}\right\|_2$ to monitor optimization health. The method demonstrates a 2.57 percentage-point improvement in classification accuracy and about a 9× reduction in run-to-run variance at 5% sampling on Fashion-MNIST, while maintaining PSNR and enabling efficient inference, outperforming efficient baselines like ISTA-Net+ and ADMM-CSNet. This physics-semantics co-design enables robust, low-bandwidth edge perception suitable for IoT/robotic deployments and highlights promising directions for domain-specific features and hardware evaluations.

Abstract

We propose PISE, a physics-informed deep ghost imaging framework for low-bandwidth edge perception. By combining adjoint operator initialization with semantic guidance, PISE improves classification accuracy by 2.57% and reduces variance by 9x at 5% sampling.

PISE: Physics-Anchored Semantically-Enhanced Deep Computational Ghost Imaging for Robust Low-Bandwidth Machine Perception

TL;DR

This work tackles robust machine perception from severely undersampled computational ghost imaging measurements, where the forward model is with and sampling rate . It proposes PISE, which combines physics-anchored initialization via and semantically guided reconstruction through a frozen VGG-based perceptual loss balanced with an MSE term, along with a gradient dynamics metric to monitor optimization health. The method demonstrates a 2.57 percentage-point improvement in classification accuracy and about a 9× reduction in run-to-run variance at 5% sampling on Fashion-MNIST, while maintaining PSNR and enabling efficient inference, outperforming efficient baselines like ISTA-Net+ and ADMM-CSNet. This physics-semantics co-design enables robust, low-bandwidth edge perception suitable for IoT/robotic deployments and highlights promising directions for domain-specific features and hardware evaluations.

Abstract

We propose PISE, a physics-informed deep ghost imaging framework for low-bandwidth edge perception. By combining adjoint operator initialization with semantic guidance, PISE improves classification accuracy by 2.57% and reduces variance by 9x at 5% sampling.
Paper Structure (14 sections, 4 equations, 3 figures, 4 tables)

This paper contains 14 sections, 4 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Forward Model
  4. PISE Framework
  5. Physics Anchor via Adjoint Proxy
  6. Semantic Enhancement via Feature-Space Loss
  7. Gradient Dynamics Metric
  8. Experimental Results
  9. Experimental Setup
  10. Comparison with Efficient Baselines at 5% Sampling
  11. Computational Efficiency
  12. Stability: Multi-Run Statistics
  13. Robustness and Error Analysis
  14. Conclusion

Figures (3)

  • Figure 1: Physically-Informed Deep CGI Workflow (Expanded View): (Left) Optical sensing (5% sampling). (Center) Physics-anchored initialization via Adjoint Operator ($A^T$) creates a coarse proxy $\mathbf{x}_{\text{init}}$. (Right) Semantically-enhanced U-Net guided by frozen VGG features ($\mathcal{L}_{\text{Perc}}$) recovers high-frequency details.
  • Figure 2: Visual comparison at $\gamma=5\%$. The adjoint proxy is severely degraded. ISTA-Net+ exhibits aliasing, U-Net (MSE) oversmooths discriminative textures, and cGAN-CS may hallucinate measurement-inconsistent details. PISE recovers semantically faithful structures without spurious artifacts. Note the recovery of high-frequency cues (e.g., shoe laces) compared to blurry MSE results.
  • Figure 3: Robustness analysis under measurement noise. The proposed PISE exhibits a flatter decay profile compared to the baseline U-Net (MSE). Notably, at high noise levels ($\sigma=0.2$), PISE remains competitive with the MSE baseline, demonstrating superior resilience.