[论文解读] Radio-Interferometric Image Reconstruction with Denoising Diffusion Restoration Models

Michel Morales, Emma Tolley+1 more|arXiv (Cornell University)|Jan 22, 2026

Radio Astronomy Observations and Technology被引用 0

一句话总结

本文提出基于 DDRM 的、数据驱动先验用于射电干涉成像重建，该先验对望远镜配置无关，在与 CLEAN 与先前扩散模型方法比较时实现了最先进的 PSNR/MSE 性能。

ABSTRACT

Reconstructing images of the radio sky from incomplete Fourier information is a key challenge in radio astronomy. In this work, we present a method for radio interferometeic image reconstruction using a data-driven prior for the radio sky based on denoising diffusion probabilistic models (DDPMs). We first train a DDPM on radio galaxy observations from the VLA FIRST survey. We create simulated VLA, EHT, and ALMA observations of radio galaxies, then use an unsupervised posterior sampling method called Denoising Diffusion Restoration Models (DDRM) to reconstruct the corresponding images, using our DDPM as a prior. Our approach is agnostic to the measured radio interferometric data and naturally incorporates the physics of the measurement process. We are able to reconstruct images with very high fidelity PSNR>60, a marked improvement over CLEAN and similar image reconstruction methods using conditional DDPMs

研究动机与目标

• Motivate improved image reconstruction for radio interferometry from incomplete Fourier data.
• Propose a data-driven prior based on denoising diffusion probabilistic models trained on radio galaxy images.
• Integrate the DDPM prior into a DDRM sampling framework to reconstruct images consistent with measurements.
• Evaluate reconstruction quality across multiple telescope configurations and compare to CLEAN.
• Explore uncertainty representation and limitations of the proposed approach.

提出的方法

• Train a 40M-parameter DDPM (U-Net with FiLM conditioning) on VLA FIRST and Radio Galaxy Zoo data to learn radio galaxy morphology.
• Formulate the DDRM sampling scheme to combine the DDPM prior with a linear forward model y = Hx + z, using an SVD-based transformation H = USV^T to operate in the singular space.
• Compute bar y = Σ^+ U^T y and bar x = V^T x to handle the measurement operator and apply conditional sampling equations (12)-(14) for different singular value regimes.
• Adapt DDRM to aperture synthesis with H = S F, using a memory-efficient SVD of S (S = I K P) and a Fourier-based forward model.
• Train with data augmentation and standard DDPM training procedures, 300k optimizer steps, AdamW optimizer, and scheduled learning rate.
• Compare DDRM reconstructions to Hogbom CLEAN using MSE, PSNR, SNR, and SRE metrics; assess sampling step trade-offs (K) and reconstruction speed.

Figure 1 : Neural Network Architecture used for the DDPM network.

实验结果

研究问题

• RQ1 Can a DDPM-based prior learned from radio galaxy morphologies serve as an effective, telescope-configuration-agnostic prior for radio interferometric image reconstruction?
• RQ2 How does DDRM perform relative to CLEAN and to conditional diffusion reconstructions across VLA, EHT, and ALMA configurations in terms of MSE/PSNR/SNR/SRE?
• RQ3 What is the impact of the number of DDRM sampling steps on reconstruction quality and computational time?
• RQ4 Can DDRM provide uncertainty quantification through reconstruction ensembles, and what are its calibration limitations?
• RQ5 How can the forward model be extended to account for W-term and A-term effects in a more realistic interferometric setting?

主要发现

• DDRM with a DDPM prior yields high-fidelity reconstructions with PSNR up to 62.9 (VLA at 1000 steps) and markedly lower MSE than CLEAN.
• DDRMs achieve better PSNR/SNR than conditional DDPM approaches on comparable uv-coverages (e.g., EHT achieved PSNR 61.7; ALMA 62.8).
• Increasing the number of sampling steps K improves MSE but increases computation time; very few steps (e.g., K=10) still yield competitive PSNR (~45–55) with fast runtimes.
• The method provides per-pixel uncertainty maps via reconstruction ensembles, though it tends to be overconfident (SRE up to 8 observed) and calibration could be improved.
• DDRMs are agnostic to the measured data and robust to different telescope configurations, though image size is limited to 150x150 due to training data constraints.

Figure 2 : A selection of radio galaxies from the test dataset.