acceptodds
Withdrawn from ICLR 2027

SALR-Former: Learning Ratio- and State-Adaptive Reconstruction for Image Compressive Sensing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image compressive sensing (ICS) reconstructs images from highly undersampled measurements. Most ICS networks are optimized for fixed sampling ratios, whereas existing arbitrary-ratio methods mainly make a shared sensing or reconstruction pipeline compatible with multiple ratios. Such compatibility does not specify how reconstruction should change during a finite unfolding process: the sampling ratio changes the measurement-constrained image subspace, while the evolving iterate determines the residual errors that remain to be corrected. We propose SALR-Former, a unified unfolding network for arbitrary-ratio ICS. Its Ratio-Stage Controller (RSC) uses the sampling ratio as a global sensing condition and combines it with the stage index, measurement residual, reconstruction feature, and cross-stage memory to control two complementary mechanisms: a bounded physics-update step size and lightweight FiLM-LoRA modulation of the learned prior. The shared prior integrates local, contextual, and frequency-domain representations, while the predicted controls specialize its behavior without ratio-specific backbones or adapter banks. Experiments on five benchmarks demonstrate strong reconstruction quality across trained and unseen continuous sampling ratios and controlled ablations further show that the sampling ratio determines the dominant reconstruction regime, whereas stage and cross-stage history refine the control trajectory.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Compressive Sensing, Deep Unfolding, Arbitrary-Ratio Reconstruction, Low-Rank Adaptation, Computational Imaging

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5rYPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SALR-Former for arbitrary-ratio image compressive sensing. A shared Ratio-Stage Controller predicts data-consistency step sizes and FiLM-LoRA modulation parameters using the sampling ratio, unfolding stage, measurement residual, reconstruction features, and recurrent memory. The reconstruction prior combines local convolution, window attention, and frequency-domain processing within a shared unfolding architecture. Evaluation covers five benchmark datasets, trained and unseen sampling ratios, measurement noise, component ablations, and inference-time control interventions.

Critical strengths
  1. The joint adaptation of the physical update and learned prior is a coherent approach to reconstruction under changing acquisition budgets. Reusing the backbone and low-rank factors across sampling ratios avoids maintaining separate ratio-specific networks, making the design practically relevant for a compact arbitrary-ratio reconstruction system.
  2. The empirical evaluation has useful breadth. The paper reports competitive results on five benchmarks and evaluates unseen sampling ratios, noise robustness, and control interventions. For example, Table 1 reports average PSNR improvements of 0.74 dB and 0.41 dB over the strongest arbitrary-ratio baselines on BSD68 and OST300, respectively. The interventions also provide an opportunity to examine the controller beyond aggregate reconstruction scores.
Critical weaknesses
  1. The central contribution of image-dependent state and history feedback is not sufficiently isolated. Table 5 reports that adding memory to the state-aware variant improves Set11 and BSD68 by 1.11 dB and 0.85 dB at a sampling ratio of 0.10. However, replacing the residual, feature, and memory descriptors with dataset-level means in Table 4 produces only 0.02–0.10 dB degradation on the other evaluated datasets. These results are not directly contradictory because the datasets and intervention protocols differ, but they do not yet provide a consistent account of the contribution of state feedback. Ratio swapping and stage shuffling demonstrate the importance of matching controls to the sensing condition and unfolding stage. They do not establish how much additional benefit comes from image-dependent feedback rather than a learned ratio-and-stage schedule. It is also unclear whether the static stage-wise baseline in Table A6 retains ratio conditioning. This distinction materially affects my assessment because state- and history-dependent control is a central part of the claimed contribution.
  2. Important ablation results are difficult to interpret without more precise configuration definitions and consistency checks. At a sampling ratio of 0.20, “w/o FiLM” in Table 2 and “LoRA only” in Table 3 appear to describe closely related configurations, yet their PSNR values differ by 0.91 dB on Set11 and 0.50 dB on BSD68. The complete-model results agree exactly between the two tables. The manuscript does not explain which architectural or training differences account for these discrepancies. Table A5 also reports unusually strong interactions: at a sampling ratio of 0.50, adapting only the step size or only the prior improves PSNR by 0.24 dB or 0.45 dB, whereas joint adaptation improves it by 3.36 dB. Strong nonlinear interactions are possible, but the interpretation depends on how the fixed-control baselines are parameterized and optimized. Precise configurations and repeated runs of the key variants are needed before these results can convincingly establish the contribution of the proposed components.
Questions for authors
  1. Are “w/o FiLM” in Table 2 and “LoRA only” in Table 3 intended to be equivalent configurations? Please specify any differences in retained modules, control targets, sensing matrices, training settings, and random seeds, and reconcile the reported results. For Table A5, please also clarify whether the fixed step sizes are independently learned per stage or shared constants.
  2. Which sampling ratios and aggregation protocol are used in Table 4? Are the masking means computed separately for each ratio and stage, or pooled across these conditions? Does the static baseline in Table A6 retain ratio conditioning? A comparison using controls averaged across images at fixed ratio–stage pairs, with the same checkpoint and evaluation data, would directly affect my assessment of the state-adaptation claim.
  3. What is the exact parameterization and range of the “bounded” step size? Is its range related to the sensing operator’s spectral norm? Appendix A.1 analyzes the unrefined gradient, whereas the implemented GRM applies a nonlinear correction. Please clarify the intended scope of this analysis and whether the GRM update is empirically observed to improve the data-fidelity term.
  4. How are the learnable sensing matrices initialized and constrained in scale? Are sensing matrices shared or retrained across the relevant ablations and competing methods? In the noise experiment, how are measurement energy and noise normalized so that comparisons correspond to comparable sensing conditions?
Reviewer_26iTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a ratio- and state-adaptive reconstruction framework called SALR-Former for image compressive sensing. A Ratio-Stage Controller predicts stage-specific update step sizes and feature modulation parameters based on the sampling ratio, reconstruction state, and cross-stage memory. The method integrates these controls into a shared deep unfolding architecture to support reconstruction at both trained and unseen sampling ratios. Experiments on multiple benchmark datasets demonstrate competitive reconstruction performance and parameter efficiency.

Critical strengths
  1. The paper considers reconstruction under varying sampling ratios and intermediate reconstruction states, which is relevant to practical CS systems where a single reconstruction model may need to operate under different measurement budgets.
  2. The framework combines ratio-aware control, stage-dependent adaptation, cross-stage memory, and feature modulation within a unified reconstruction pipeline. This provides a clear overall structure for studying adaptive reconstruction.
  3. The reported model size is relatively small compared with some competing methods. This is a practical advantage, although the computational cost and inference latency should be considered alongside the parameter count.
Critical weaknesses

1.The idea of generating adaptive parameters to control iterative reconstruction has been explored in existing deep unfolding methods. For example, DPUNet (Neurocomputing2022) and D3C2-Net (TCSVT2024). The novelty of SALR-Former therefore remains unclear. The authors should clarify its differences from prior approaches and demonstrate the specific benefits of jointly adapting to the sampling ratio, reconstruction stage, and intermediate reconstruction state. 2.The motivation for adopting the bilateral structured sensing scheme is insufficiently justified, as a single learnable sensing matrix can also support multiple sampling ratios by selecting different numbers of rows. 3.State-masking experiments show relatively small performance drops on several datasets, raising questions about the independent contribution of reconstruction-state information and cross-stage memory. More controlled ablations are needed to validate the effectiveness of these components.

Questions for authors

1.How does SALR-Former differ from existing adaptive parameter generation methods, such as DPUNet (Neurocomputing2022) and D3C2-Net (TCSVT2024). 2.Why do the authors choose the bilateral structured sensing scheme instead of a single learnable sensing matrix with row-wise truncation? Does the proposed design offer clear advantages in terms of parameter efficiency, computational cost, or reconstruction performance? 3.Given the relatively small performance changes in some state-masking experiments, what evidence supports the claim that reconstruction-state adaptation and cross-stage memory provide a substantial benefit beyond sampling-ratio conditioning?

Reviewer_pCpJOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SALR-Former, a unified deep unfolding framework for arbitrary-ratio image compressive sensing. Its key component, the Ratio-Stage Controller (RSC), dynamically adjusts the physics-guided update step size and the learned reconstruction prior according to the sampling ratio, unfolding stage, reconstruction residual, features, and cross-stage memory. FiLM-LoRA modulation is employed to adapt the shared reconstruction network. Experiments on five benchmark datasets report competitive reconstruction performance across trained and unseen sampling ratios.

Critical strengths
  1. Interesting adaptive unfolding framework. The proposed RSC jointly modulates the physics-guided update and learned prior, providing a potentially useful mechanism for adapting reconstruction to different sensing conditions and unfolding stages.

  2. Extensive experimental evaluation. The paper reports results across multiple datasets and sampling ratios, including unseen ratios. The ablation studies and control-intervention experiments provide useful insights into the proposed adaptive mechanisms.

Critical weaknesses
  1. Limited novelty over existing arbitrary-ratio reconstruction methods. The third contribution emphasizes single-model reconstruction across trained and unseen sampling ratios. However, ProxUnroll (CVPR 2025) has already demonstrated this capability without retraining or fine-tuning. The main novelty therefore lies in the proposed RSC, whose advantages over existing arbitrary-ratio unfolding frameworks need to be more convincingly established.

  2. Significant discrepancies in baseline reconstruction performance. Several baseline results are substantially lower than previously published values. For example, the original ProxUnroll paper reports 23.51 dB at a 1% sampling ratio on CBSD68, whereas this submission reports only 20.27 dB on BSD68. Its reported average PSNR is also 1.52 dB lower than the original publication. Similarly, the reported SIB-ACS results are 0.31–0.77 dB lower at several sampling ratios. Published reference results for DGUNet+ and HATNet also suggest substantial discrepancies on Set11. Although evaluation protocols may differ, these differences require careful verification and explanation, as they directly affect the claimed superiority of SALR-Former.

  3. Insufficient evidence for state-adaptive reconstruction. Although state-dependent adaptation is a central motivation, Table 4 shows that masking reconstruction-state information reduces PSNR by only 0.02–0.10 dB, much less than ratio swapping or stage shuffling. Meanwhile, Table 5 reports considerably larger improvements from the complete RSC. The relationship between these observations remains unclear, and the independent contribution of state-aware conditioning is not sufficiently established.

  4. Limited theoretical justification for the learned updates. Appendix A analyzes the spectral behavior of standard gradient descent to motivate adaptive step sizes. However, the actual unfolding network incorporates nonlinear gradient refinement and learned prior modulation. The theoretical analysis therefore does not directly establish the stability or convergence of the implemented reconstruction process.

Questions for authors
  1. How does the proposed framework differ from ProxUnroll beyond the RSC mechanism?

  2. Why are the reported results for ProxUnroll and several other baselines substantially lower than previously published values? Can the authors verify and explain these discrepancies?

  3. How do the authors explain the negligible effect of state masking in Table 4 compared with the substantial gains from Full RSC in Table 5?

  4. Does the theoretical analysis extend to the actual nonlinear unfolding updates?

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

Related papers

Discussion (0)

Sign in to comment.