acceptodds
Under review as a conference paper at ICLR 2027

Correcting Spectra Outside the Backbone: A Model-Agnostic Rectifier for Hyperspectral Image Super-Resolution

Abstract

Hyperspectral image super-resolution (HSI-SR) aims to recover spatial detail while preserving the spectral shape on which quantitative analysis relies. Recent HSI-SR methods, from repurposed RGB super-resolution backbones to dedicated spectral-spatial architectures, have greatly improved spatial reconstruction. However, overlooking the compact spectral structure of hyperspectral data leaves residual spectral errors, while binding the spectral treatment to each architecture forces it to be rebuilt for every new backbone. Yet the low-dimensional structure of spectra belongs to the data, not to any backbone. Backbones differ in the errors they leave, but not in the structure of the true spectra. One rectifier design can therefore serve any backbone. Building on this insight, we propose the Spectral Rectification Super-Resolution Network (SR²-Net), a model-agnostic rectifier that needs nothing from the backbone but its output, leaves its internal architecture untouched, and is trained per backbone. SR²-Net follows an enhance-then-rectify pipeline in which Hierarchical Spectral-Spatial Synergy Attention (H-S³A) reinforces cross-band interactions, while Mode-Constrained Rectification (MCR) confines the correction to a learned compact spectral subspace. A degradation-consistency constraint further ties the output to the observed low-resolution input. Experiments with five backbones spanning CNN, Transformer, and diffusion families show that one fixed configuration improves spectral fidelity in every reported setting while preserving or improving spatial quality. Averaged over thirty in-domain settings, SR²-Net removes 22.6% of the residual spectral error at a backbone-independent cost of 0.048M parameters.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
hyperspectral image super-resolution, model-agnostic rectification

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_6NqgOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SR²-Net, a model-agnostic spectral rectifier for hyperspectral image super-resolution. It combines hierarchical spectral-spatial attention, low-rank residual correction, and degradation consistency to improve spectral fidelity without modifying the backbone architecture. Experiments across five backbones demonstrate consistent improvements in spatial and spectral reconstruction quality.

Critical strengths
  1. The paper addresses residual spectral distortions in existing HSI-SR models through a flexible output-side correction strategy.
  2. The proposed H-S³A and MCR modules combine spectral-spatial interactions with low-dimensional residual constraints to improve spectral fidelity.
Critical weaknesses
  1. Limited methodological novelty. H-S³A mainly combines existing spectral grouping, attention, and channel-shuffling operations, while MCR adopts conventional low-dimensional projection and residual correction. The authors should better distinguish the proposed method from existing low-rank spectral modeling and two-stage refinement approaches.
  2. Insufficient justification of the low-rank correction assumption. The low-dimensional structure of ground-truth spectra does not necessarily imply that backbone-dependent residual errors are also low-rank. Moreover, MCR constrains only the predicted corrections, not the final reconstructed spectra. Further analysis is needed to substantiate this central assumption.
Questions for authors
  1. Since SR²-Net requires separate training for each backbone, to what extent is it truly model-agnostic? Can a trained rectifier transfer across different backbones without retraining?
  2. Table A11 examines the scenario of training only the SR²-Net while freezing the backbone. Would continuing to train the backbone for the same number of epochs as the SR²-Net yield better results? Or is the SR²-Net unnecessary under this configuration?
  3. Can the authors analyze the singular-value distributions of residual spectral errors and the similarity of learned spectral modes across different backbones?
  4. The comparisons with parameter-matched learned rectifiers are mainly conducted on SwinIR at ×4. Do the advantages persist across other architectures and scaling factors?
  5. How does the method handle different numbers of spectral bands, and does a fixed spectral grouping offer any advantages? Also, how does it perform in terms of generalization?
Reviewer_TzryOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SR²-Net, a lightweight rectifier attached to the output of a hyperspectral SR backbone. It has three parts:

  • H-S3A: cross-band interaction enhancement module with "spectral grouping, trilateral synergy attention, and adjacent spectra shuffle".

  • MCR: a rank-8 bottleneck whose correction is added to the backbone’s output. It models the rectification effect the way "hyperspectral data are modeled, as r = 8 learned spectral modes shared by all pixels"

  • A degradation-consistency loss: the output, downsampled with bicubic interpolation, should match the low-resolution input.

The rectifier is trained jointly with each of five backbones (RCAN, SSPSR, SwinIR, ESSAformer, DDS2M) on ARAD-1K and CAVE at ×2, ×4 and ×8. It improves spectral fidelity in every reported setting, transfers to CAVE and ICVL, and works on frozen backbones.

Critical strengths

S1. Very extensive experimental setting. Five backbones from three families, three scales, in-domain and cross-domain tests, a frozen-backbone variant, and test-time degradation mismatch are reported.

S2. The design is very light-weight and can be easily plugged into any existing models.

Critical weaknesses

W1.The model-agnostic claim can be overstated and misleading. The rectifier is essentially retrained for every backbone. It is the module design, not trained weights, that carry over. It is a little bit weird to call this model-agnostic -- that sounds like calling a linear layer "model agnostic".

Model-agnostic post-hoc modules for SR are not new. Explorable Super Resolution (Bahat & Michaeli, CVPR 2020) presented an architecture that "wraps any SR network and guarantees analytically that its outputs exactly match the LR input when downsampled". While I am not an expert in this, it theoretically sounds like what this paper's proposed module is doing. It should be cited and compared, if possible.

Questions for authors

Q1. Have you compared with other rectifiers that control the difference between SR and LR spectra?

Reviewer_fqbtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This manuscript addresses the problem of hyperspectral super-resolution by designing a rectifier which acts as a plug-and-play post-processing module that corrects the residual spectral error thus maintaining the spectral fidelity of the hyperspectral image.

Critical strengths
  1. The insights that low-dimensional spectra structure exists in the hyperspectral data itself instead of any network backbone is interesting and potentially inspiring. The observation that spectral errors rather than true spectra matter is also insightful.
  2. The experiments are well-designed and sufficiently comprehensive. Both the quantitative and visualisation results support well effectiveness of the proposed method.
Critical weaknesses
  1. The overall narrative could be presented in a clearer and more straightforward manner, as the current organization makes the manuscript somewhat difficult to follow.
  2. In Figure 3, A1, A2 and A3, 'ESSAFOEMER' should be spelt as 'ESSAFORMER'
Questions for authors

It can be observed from Table 1 that mature RGB-based super-resolution methods, such as RCAN and SwinIR, achieve the best or second-best performance on most of the reported evaluation metrics. This observation raises an important question regarding the source of these results. One possible explanation is that the networks are overfitting to the characteristics of the benchmark datasets; another is that the datasets themselves may be constrained by relatively old sensing hardware (such as CAVE, released on 2007) and therefore may not fully reflect the characteristics of more recent hyperspectral imagery. This leads to two concerns:

  1. Regarding the network design: Given that RGB-based architectures already achieve highly competitive, and in many cases superior, results, to what extent is an HSI-specific architecture necessary? The authors are encouraged to clarify whether simply adapting mature RGB-based networks is sufficient for this task, and what additional benefit is provided by explicitly modeling hyperspectral characteristics.

  2. Regarding the datasets: Could the authors consider evaluating the proposed method on more recent hyperspectral datasets acquired with newer sensors? Such experiments would help verify whether the conclusions generalize beyond the current benchmarks and are not overly dependent on potentially outdated sensing conditions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.