- 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
- 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.
- 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
- 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.
- 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
- 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.
- 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.
- 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.
- 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?