CVPR2025 Image Restoration radio interferometry image reconstruction super-resolution deep learning strong gravitational lensing high dynamic range

POLISH'ing the Sky: Wide-Field and High-Dynamic Range Interferometric Image Reconstruction with Application to Strong Lens Discovery¶

Conference: CVPR2025
arXiv: 2603.09162
Code: To be confirmed
Area: Image Restoration
Keywords: radio interferometry, image reconstruction, super-resolution, deep learning, strong gravitational lensing, high dynamic range

TL;DR¶

Based on the POLISH framework, POLISH+/++ is proposed with two key improvements—patch-wise training + stitched inference and arcsinh non-linear transformation—enabling deep learning methods to handle wide-field ($12,960 \times 12,960$ pixels) and high-dynamic-range ($\sim 10^6$) radio interferometric imaging for the first time, while demonstrating a $10\times$ potential increase in strong gravitational lensing discovery through super-resolution.

Background & Motivation¶

Computational Challenges of Radio Interferometric Imaging: The upcoming Deep Synoptic Array (DSA) will feature 1650 antennas with a raw data throughput exceeding 80 Tb/s, which traditional methods cannot process in real time.

Limitations of Prior DL Methods: As summarized in Table 1, existing DL methods have only been tested on small images ($<1000$ pixels) and low dynamic ranges ($<10^3$), which is far from practical deployment requirements.

Dimensional Bottleneck of Wide-Field: The field of view of DSA is approximately 10 square degrees, with images exceeding $10,000 \times 10,000$ pixels. A single float32 input requires about 400MB of GPU memory, and feature map storage is even larger.

High Dynamic Range Issue: The ratio between bright and faint sources in a wide field can reach $10^4$–$10^6$. Direct training causes the network to bias toward bright sources while ignoring faint ones.

Model Mismatch: In actual observations, ionospheric effects, antenna pointing errors, and other factors cause the PSF to deviate from assumptions, a problem that is rarely considered in existing works.

Potential for Strong Gravitational Lensing Discovery: DSA is expected to discover $10^4$–$10^5$ strong lenses, but traditional CLEAN methods are limited by the PSF resolution. Super-resolution has the potential to increase the discovery rate by an order of magnitude.

Method¶

Overall Architecture¶

Based on POLISH (an end-to-end CNN mapping dirty images to clean images), two key improvements targeting wide-field and high-dynamic-range imaging are proposed: - POLISH+: Patch-wise training + stitched inference - POLISH++: Patch-wise training + arcsinh transformation

Key Designs¶

1. Patch-Wise Processing - Slice the $12,960 \times 12,960$ image into $J$ non-overlapping patches ($324 \times 324$ pixels). - Each dirty-clean training pair is divided into $J$ patch training pairs; 18 training images yield 28,800 training patches. - Key Insight: The dirty image in a patch contains PSF sidelobe artifacts originating from bright sources in adjacent patches (cross-patch contamination). This is significantly different from simply convolving the patch with the PSF, implying that the network needs to implicitly handle non-local effects. - During inference, the patch predictions are stitched back into the full-field-of-view image.

2. Arcsinh Non-linear Transformation - Transformation function: $\text{AsinhStretch}(x; a) = \frac{\text{arcsinh}(x/a)}{\text{arcsinh}(1/a)}$ - Compresses pixel values spanning multiple orders of magnitude into the same order, reducing the dynamic range by more than an order of magnitude. - Superior to gamma encoding: Can handle both positive and negative values (dirty images can contain negative values). - Both training and inference are conducted in the transformed space, and the inverse transformation $\text{AsinhStretch}^{-1}$ is applied after inference to restore the original scale. - Hyperparameters $a_{\text{dirty}}$ and $a_{\text{true}}$ control the compression intensity of the input and target, respectively.

Loss & Training¶

The $\ell_1$ loss is computed in the arcsinh-transformed space: $$\theta^* = \arg\min_\theta \frac{1}{NJ} \sum_{i,j} \|G_\theta(\text{AsinhStretch}(\mathbf{I}^{[j]}_{\text{dirty}}; a_{\text{dirty}})) - \text{AsinhStretch}(\mathbf{I}^{[j]}_{\text{true}}; a_{\text{true}})\|_1$$

Key Experimental Results¶

Main Results: Source Detection Accuracy¶

Method	Precision↑	Recall↑	F₁↑
CLEAN	0.3612	0.2220	0.2750
POLISH	0.5560	0.4612	0.5042
POLISH+	0.8744	0.5751	0.6938
POLISH++	0.8433	0.6142	0.7107

POLISH++ improves $F_1$ score by 158% compared to CLEAN ($0.2750 \rightarrow 0.7107$), and by 41% compared to POLISH.

Shape and Flux Estimation Accuracy (RMSE)¶

Method	Major Axis FWHM (″)↓	Minor Axis FWHM (″)↓	Flux (Jy/pix)↓
CLEAN	1.0046	0.7862	$3.26 \times 10^{-4}$
POLISH++	0.4654	0.2056	$3.17 \times 10^{-3}$

Shape estimation accuracy is significantly improved, whereas CLEAN remains superior in flux estimation (as CLEAN preserves absolute flux calibration).

Strong Gravitational Lensing Discovery¶

The super-resolution of POLISH/++ enables the CNN lens finder to recover lenses with Einstein radii close to the PSF scale.
Compared to traditional CLEAN (requiring separation $>3\times$ PSF), POLISH reduces the lower limit of detection to $\sim1\times$ PSF.
The number of discoverable lenses in the DSA survey is increased by approximately 10 times.

Model Robustness and Adaptability¶

Trained on ideal PSFs and tested on PSF distortions $\gamma \in [0,30]$: Visual reconstructions remain stable (with a predictable decrease in PSNR).
Fine-tuning to a new PSF distribution requires only 11 epochs (compared to 57 epochs for training from scratch), achieving a $5\times$ speedup.

Key Findings¶

Patch-wise training not only overcomes memory limits but also implicitly learns to handle cross-patch PSF sidelobe contamination.
The arcsinh transformation increases the recall of POLISH++ by 4% on low-SNR sources and improves the $F_1$ score.
Super-resolution breaks the PSF diffraction limit: POLISH++ can accurately estimate sources with angular sizes much smaller than the PSF width ($\approx 3.3''$).

Highlights & Insights¶

Application-Driven Engineering Innovation: Two seemingly simple improvements (patching and non-linear transformation) render deep learning methods applicable to real-world radio astronomy scales for the first time.
Interesting Phenomenon of Cross-Patch Contamination: Patched dirty images contain non-local artifacts from neighboring patches, yet deep learning methods can implicitly learn to handle this effect. While this theoretically should not hold (as a local forward model does not exist), experiments prove its effectiveness.
Impact of Super-Resolution on Scientific Discovery: Rather than just improving image quality, this directly translates to an order-of-magnitude increase in the number of strong gravitational lensing discoveries.
Efficient Adaptation via Fine-Tuning: The pre-training + fine-tuning strategy enables the model to rapidly adapt to different observational conditions, which is crucial for practical deployment.

Limitations & Future Work¶

Inaccurate Flux Estimation: The non-linear nature of learning-based methods results in a higher RMSE for flux estimation compared to CLEAN, lacking an explicit flux calibration mechanism.
Simple Astronomical Model Assumptions: The training utilizes T-RECS simulations (Gaussian/Sérsic profiles), and generalization to more complex morphologies (e.g., radio jets, extended structures) remains unverified.
Pixel Scale Constraints: The $\sim 1''$ pixel scale poses a fundamental limitation at extremely small separation scales.
Small Training Dataset: Training with only 18 images (though generating 28,800 patches through slice processing) may limit generalization capabilities.

vs CLEAN: CLEAN is limited by the PSF resolution and cannot perform super-resolution, but it preserves flux calibration. POLISH++ exhibits strong super-resolution capability but yields less accurate flux estimation.
vs R2D2: R2D2 was tested on $512^2$ images, whereas POLISH++ scales to $12,960^2$, representing a $600\times+$ increase in scale.
vs RML: Optimization-based methods offer high quality but are computationally expensive, making them unsuitable for DSA real-time processing demands.
Insight: Simple data transformations (such as arcsinh) can be more important than complex network designs in high-dynamic-range scenarios. The pre-training + fine-tuning paradigm is highly suitable for astronomical applications requiring adaptation to varying observational conditions.

Rating¶

Novelty: ⭐⭐⭐ — The patch-wise training and non-linear transformation methods themselves are relatively simple, but their first large-scale application in radio astronomy scenarios is significant.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Extremely thorough, with comprehensive evaluations spanning source detection, shape estimation, lens discovery, robustness, and adaptability.
Writing Quality: ⭐⭐⭐⭐ — The background introduction is clear, the experimental design is rigorous, and the analysis of astronomical domain knowledge is in-depth.
Value: ⭐⭐⭐⭐ — Enables DL methods to be deployed at the practical scale of next-generation radio telescopes for the first time, offering significant practical value to the radio astronomy community.