POLISH'ing the Sky: Wide-Field and High-Dynamic Range Interferometric Image Reconstruction¶

Conference: CVPR 2026
arXiv: 2603.09162
Code: None (extension of the POLISH framework)
Area: Radio Astronomical Image Reconstruction / Image Deconvolution
Keywords: Radio interferometric imaging, deep learning deconvolution, super-resolution, high dynamic range, strong gravitational lensing

TL;DR¶

Building upon the POLISH framework, this work proposes POLISH+ and POLISH++, which employ a patch-based training-and-stitching strategy and an arcsinh-based nonlinear transform to achieve radio interferometric image reconstruction and super-resolution under wide-field (12,960×12,960 pixels) and high-dynamic-range ($\sim 10^6$) conditions. The paper also presents the first demonstration that deep learning methods can super-resolve strong gravitational lens systems.

Background & Motivation¶

Radio interferometric imaging achieves high-resolution astronomical imaging by synthesizing large apertures from antenna arrays, with image deconvolution as its core problem. The forthcoming DSA-2000 (1,650 antennas) will produce images exceeding $10,000 \times 10,000$ pixels with a dynamic range of $\sim 10^6$ at data throughput rates exceeding 80 Tb/s.

Limitations of existing methods:

CLEAN: The standard approach, but resolution is limited to the PSF scale with no super-resolution capability; poorly suited to recovering complex source morphologies.

RML optimization methods: High computational cost of iterative solving makes them unsuitable for DSA's real-time requirements.

Existing deep learning methods: - Small image sizes (< 1,024 pixels) and low dynamic ranges (< $10^3$) - Tested only on simple Gaussian sources, without addressing complex celestial morphologies (e.g., strong lenses) - PSF mismatch (calibration errors) not considered

Core challenges: (1) extremely high dimensionality imposed by wide fields of view; (2) extremely high dynamic ranges of real skies.

Method¶

Overall Architecture¶

The POLISH family employs an end-to-end CNN (based on the WDSR architecture) that directly learns the mapping from low-resolution dirty images to high-resolution clean sky images. This work extends that framework with: - POLISH+: Patch-based training and inference - POLISH++: POLISH+ augmented with an arcsinh nonlinear transform

Forward model: $I_{\text{dirty}} = [k * (I_{\text{true}} + n)]{\downarrow_s}$, where $k$ is the PSF and $s=2$ is the downsampling factor.

Key Designs¶

Patch-wise Processing: The full 12,960×12,960 field-of-view image is divided into $J$ non-overlapping 324×324 patches, each forming an individual training pair. Key insight:
- Patches cropped from full-field dirty images contain cross-patch contamination (PSF sidelobe artifacts from bright sources in neighboring patches)
- This is fundamentally different from training directly on small images — the network must learn to handle these non-local artifacts
- 18 full-field images → 28,800 training pairs (containing 6 million detectable galaxies)
- At inference, each patch is predicted independently and then stitched back into the full field

Design Motivation: GPU memory cannot accommodate training on complete images of $10^8$ pixels; patch-wise processing makes large-scale training feasible.

Arcsinh Nonlinear Transform (AsinhStretch): Addresses the extreme dynamic range of $10^4 \sim 10^6$: $$\text{AsinhStretch}(x; a) = \frac{\operatorname{arcsinh}(x/a)}{\operatorname{arcsinh}(1/a)}$$
- Logarithmic in form: compresses pixel values spanning multiple orders of magnitude into a comparable range
- Handles positive and negative values (dirty images may contain negative values), making it appropriate for interferometric imaging
- Training loss is computed in the transformed space: $$\theta^* = \arg\min_\theta \frac{1}{NJ}\sum_{i,j} \|\text{G}_\theta(\text{AsinhStretch}(I_{\text{dirty}}^{[j]}; a_d)) - \text{AsinhStretch}(I_{\text{true}}^{[j]}; a_t)\|_1$$
- At inference, original intensity scales are recovered via the inverse transform $\text{AsinhStretch}^{-1}$

Design Motivation: When training directly in the original intensity space, the $\ell_1$ loss is dominated by a small number of bright sources, leading to poor recovery of faint sources.

Model Robustness and Adaptability:
- Robustness: A model trained only on ideal PSFs maintains visual consistency when faced with randomly perturbed PSFs ($\gamma \in [0, 30]$)
- Adaptability: Fine-tuning converges more than 5× faster than training from scratch (11 vs. 57 epochs), enabling rapid adaptation to different observing conditions

Loss & Training¶

Loss: $\ell_1$ loss (computed in the AsinhStretch-transformed space)
Optimizer: Adam, lr 0.0001, batch size 12
$a_{\text{dirty}} = a_{\text{true}} = 0.1$
Training data: T-RECS sky simulations, 18 training images + 5 test images
Noise: Gaussian noise $\sigma = 1\,\mu\text{Jy}$

Key Experimental Results¶

Main Results¶

Method	Precision↑	Recall↑	F1↑	Major-axis FWHM RMSE↓	Minor-axis FWHM RMSE↓
CLEAN	0.3612	0.2220	0.2750	1.0046″	0.7862″
POLISH	0.5560	0.4612	0.5042	0.9642″	0.3219″
POLISH+	0.8744	0.5751	0.6938	0.4335″	0.1889″
POLISH++	0.8433	0.6142	0.7107	0.4654″	0.2056″

Note: POLISH++ improves F1 by 159% over CLEAN, with more than 2× improvement in shape estimation accuracy.

Ablation Study¶

Configuration	Key Metric	Note
POLISH (full-image training)	F1=0.5042	Baseline
POLISH+ (patch-based training)	F1=0.6938	Patch strategy yields large gains
POLISH++ (patch + arcsinh)	F1=0.7107	Nonlinear transform further improves recall by +4%
PSF perturbation γ=0→30	PSNR drops but visually consistent	Robust to calibration errors
Fine-tuning vs. training from scratch	11 vs. 57 epochs	5× speedup

Key Findings¶

Super-resolution capability: POLISH++ accurately estimates galaxy shape parameters below the PSF scale (≈3.3″), where CLEAN fails entirely
Strong lens discovery: A lens-finding CNN trained on POLISH++ super-resolved images can lower the discoverable lens threshold from 3× the PSF resolution limit to near the PSF scale, yielding an approximately 10× increase in the expected strong lens yield from DSA surveys
Dynamic range: POLISH++ successfully handles a dynamic range of $\sim 10^6$, three orders of magnitude higher than existing DL methods (< $10^3$)
CLEAN still outperforms POLISH in flux estimation (model-based methods retain better absolute flux calibration)

Highlights & Insights¶

Deployment-oriented design: Rather than optimizing benchmark PSNR on small images, the method is designed for DSA's practical requirements (12,960×12,960 pixels, $10^6$ dynamic range)
Discovery of cross-patch contamination: Dirty image patches contain PSF sidelobe artifacts from neighboring bright sources — a unique domain-specific insight
Scientific application value: Super-resolution directly enables strong gravitational lens discovery, increasing DSA's lens yield by 10×
Honest limitation analysis: The paper explicitly acknowledges that CLEAN remains superior for flux estimation, and that DL methods lack an explicit flux calibration mechanism
From robustness to adaptability: The paper not only validates robustness under PSF mismatch but also demonstrates rapid fine-tuning adaptability

Limitations & Future Work¶

Flux estimation accuracy is inferior to CLEAN; no explicit flux calibration mechanism is present
Operations are confined to the image plane (not the visibility domain), potentially losing phase information
Training data are based on T-RECS simulations, which may not fully represent real skies
Patch stitching may introduce discontinuities at patch boundaries (not discussed in detail)
Future directions: flux post-processing calibration, end-to-end visibility-domain methods, simulation of more complex celestial morphologies

POLISH (Connor et al. 2022): The baseline method of this work, tested only on 2,048-pixel images with a dynamic range of $\sim 10^2$
R2D2: An unrolled network approach supporting 512-pixel images and a dynamic range of $5 \times 10^5$
CLEAN: The standard method in radio astronomy, limited by PSF resolution
Insight: The "killer application" of deep learning deconvolution in astronomical imaging is the scientific discovery capability enabled by super-resolution

Rating¶

Novelty: ⭐⭐⭐ Technically an engineering improvement over existing methods (patch-based training + nonlinear transform); the core contribution lies in engineering scale and application impact
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Covers source detection, shape estimation, flux estimation, strong lens discovery, and PSF robustness/adaptability
Writing Quality: ⭐⭐⭐⭐ Seamless integration of astronomical background and DL methodology; problem formulation is clear
Value: ⭐⭐⭐⭐ High practical value for DSA deployment; the 10× improvement in strong lens yield constitutes an important scientific contribution