Image Restoration & Super-Resolution

Cross-Domain Robustness Audit of Image Restoration Models Under Physical Sensor and Optical Degradations

Effect to explainOpen

Weak candidate · 1/3 runs3 papers report this67% from 2025+

Generated automatically from the limitations stated in 3 papers (ICML, ICCV, CVPR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.

The problem

Image restoration and super-resolution models are primarily developed and evaluated under synthetic degradation models, including bicubic downsampling, Wald-protocol Gaussian blur, and simplified sensor calibrations. Consequently, it remains unknown how well these methods transfer to actual physical captures, such as high-magnification (8×/16×) microscopy and real-world non-uniform sensor acquisitions. Practitioners deploying models to physical instruments cannot anticipate whether synthetic-trained models will retain their reported performance or fail catastrophically. Without a systematic cross-setting robustness evaluation, the domain-gap penalty between synthetic simulations and real optical sensors remains unquantified.

Why it matters

Provides the first standardized measurement of real-sensor generalization gaps across standard restoration architectures. Enables researchers to identify precisely which physical degradation factors require domain-specific modeling versus standard synthetic training.

Ways to approach it

Prior-work checks are free with an account. Results someone already ran are shown to everyone.

1
Multi-domain benchmark evaluation: Assemble existing open real-capture datasets across domains (including physical microscope paired acquisitions, real sensor captures, and hyperspectral cubes) and evaluate representative restoration and blind SR models (e.g., Real-ESRGAN, DASR, all-in-one backbones) trained purely on standard synthetic pipelines, measuring PSNR, SSIM, and reference-free quality metrics.
2
Synthetic-to-physical degradation sensitivity profiling: Perturb standard degradation generators (varying noise non-uniformity, optical blur profiles, and subsampling factors) to measure the exact error tolerance thresholds beyond which synthetic-trained models break down relative to real sensor baselines.

Have a different approach?

Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.

Why it might fail

Difficulty in securing verified paired ground-truth data for real microscope and physical sensor setups could limit evaluation to no-reference metrics, weakening quantitative conclusions. Alternatively, standard blind restoration methods might already generalize adequately across these settings, leaving little performance gap to report.

Evidence

Each paper's own statement of the limitation, verbatim.

SSDCN: Spatial-Spectral Dual-Clustering-based Network for Hyperspectral Image Super-resolutionICML 2026
Evaluation limited to three small datasets with Wald-protocol synthetic Gaussian-blur degradations (187/92/80 cubes, 3:1 train/test); real sensor degradation is not tested
Physical Degradation Model-Guided Interferometric Hyperspectral Reconstruction with Unfolding TransformerICCV 2025
Training relies entirely on synthetic data generated from the simplified degradation model; residual mismatch between simulation and real sensor behavior would directly limit real-world performance, and only uniform-light calibration scenes are used for real testing
Learning Large-Factor EM Image Super-Resolution with Generative PriorsCVPR 2024
Performance depends on synthetic bicubic-downsampled degradation, so real microscope degradation at 8×/16× may not match training assumptions (Real-ESRGAN comparison partially addresses robustness, but degradation realism is unverified on real paired data)