acceptodds
Withdrawn from ICLR 2027

Action-Regulated State Spaces for All-in-One Endoscopic Image Restoration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

All-in-one medical image restoration faces a fundamental spatial action-allocation problem: degradations vary widely in location, extent, and severity. We identify a previously overlooked failure mode of state-space restorers under such spatial heterogeneity. Because corrective evidence is recurrently written into and propagated through shared state dynamics, an action appropriate for one region may cross its intended boundary and inadvertently alter intact tissue. We term this phenomenon restoration-action spillover. To address this issue, we proppse BARS-MambaIR, a blind action-regulated state-space model that predicts dense demand, scope, and magnitude fields to encode where, how far, and how strongly restoration should act. These fields regulate state writing, propagation decay, boundary-conditioned attenuation, and local-to-global action routing, while also gating the output residual. An action-allocation consistency objective learns these fields directly from paired supervision, without requiring degradation labels or region annotations, to suppress intact-region distortion and restoration spillover. Controlled footprint–severity sweeps, localized state interventions, and field-swapping experiments further reveal how spatially regulated state propagation limits restoration-action spillover.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
All-in-One Image Restoration, State-Space Models, Spatially Adaptive Restoration, Restoration-Action Spillover, Mamba

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

This paper focuses on 'restoration-action spillover' in endoscopic image restoration. The authors propose BARS-MambaIR, which modulates the scope of state propagation and the residual output by predicting three types of spatial control signals: demand, scope and magnitude. Experiments were conducted on the SurgClean and SegSTRONG-C datasets, and their performance was analysed using spillover and some other metrics.

Critical strengths
  1. The problem formulation is somewhat novel. The paper analyses the impact of restoration operations on non-degraded regions from the perspective of ‘restoration-action spillover’, and attempts to explicitly model the location, extent and intensity of the restoration effects.

  2. The experimental analysis is quite comprehensive. In addition to the standard PSNR, SSIM and LPIPS metrics, the authors have devised spillover, boundary-distance, intervention and routing analyses to elucidate the model’s spatial control behaviour.

Critical weaknesses
  1. The concept of ‘restoration-action spillover’ is of limited novelty. This issue is closely related to phenomena such as over-restoration and spatial leakage in existing image restoration methods. Furthermore, the method design is relatively complex and exhibits significant engineering stacking. The model introduces multiple components, including demand, scope, magnitude, boundary barrier, multi-range routing, retention decay and action-conditioned prompts; consequently, its core contribution lacks focus.

  2. Some key hyperparameters and control rules are rather ad hoc. For example, parameters related to different propagation bases are set manually, but there is a lack of sufficient sensitivity analysis or theoretical justification.

  3. The core ‘Spill’ metric relies on a degradation-support map constructed by the authors. This support is derived from the input-reference residual; the reliability and generalisability of this metric still require further validation.

  4. The mechanism analysis is of limited persuasiveness. The paper primarily illustrates the roles of various control variables through internal routing statistics and manual intervention; however, these experiments appear more like ‘self-validation’ and are insufficient to demonstrate that they correspond to a genuine and generalisable restoration mechanism.

  5. Lack of more thorough validation of the backbone. The method relies almost entirely on MambaIRv2 and has not been validated on other state-space restoration backbones; consequently, its generalisability is difficult to guarantee.

  6. Insufficient comparison with more direct spatial gating baselines. Given that the core idea is to ‘limit the scope of restoration’, the method should have been compared with more direct approaches such as simple spatial mask gating, attention gating, residual masking and dynamic convolution; otherwise, it is difficult to demonstrate that such complex state regulation is necessary.

  7. A lack of statistical significance analysis. Many of the improvements in the main table are actually very small; for example, the PSNR for SurgClean defog only increased from 19.49 to 19.55, yet there is no significance test to indicate whether these improvements are stable.

Questions for authors
  1. Could the authors further clarify the fundamental distinction between ‘restoration-action spillover’ and phenomena such as over-restoration, spatial leakage or unintended modifications to intact regions, as observed in existing image restoration research?

  2. What evidence from the current experiments demonstrates that such a complex, multi-component state-regulation design is necessary, rather than similar benefits being achievable through simpler spatial gating mechanisms?

  3. How were the manually set propagation parameters in the paper selected? To what extent are the results sensitive to these parameters?

Reviewer_X3oqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work aims to solve the task of blind all-in-one restoration of endoscopic images, while focusing on fixing the problem of state-space-based (Mamba-like) restoration models. They claim that unwanted modifications and corrections affect the intact tissue region (restoration-action spillover), due to the nature of the state-space models. The solve this by formulating the restoration problem as controlled action transport scheme, adding a controller which predicts dense demand, scope and magnitude. Experimental results on benchmark datasets show improved performance over baseline, and competitiveness among conventional methods.

Critical strengths
  • The proposed method seems to deliver some benchmark gains under a common all-in-one protocol.
  • The authors' attempt to propose a new spill metric, that could be adapted to evaluating other restoration methods, is somewhat promising to some extent.
  • The motivation and proposed solution of the authors are clear and straightforward.
Critical weaknesses
  • The proposed method's core contribution of reduced spillover, claimed by the authors, seems to be based on a metric that were defined by the authors themselves.
  • The authors optimize their method based on a metric that was defined by the authors. This was not performed on other methods, which would be an unfair comparison.
  • The correlation of spillover metric and final visual quality is not analyzed in-depth. The practical significance of the proposed spillover metric remains unclear.
  • Given the results in Table 1, quantitative performance gains over MambaIRv2 does not seem to be statistically meaningful on many tasks and metrics.
  • Given increased capacity of the framework with additional modules, comparison under same parameter count, FLOPs, and memory usage should be performed for fairness.
  • Since several new mechanisms are introduced simultaneously, it is difficult to determine which component is primarily responsible for the observed performance gains.
Questions for authors

Please refer to the weaknesses section.

Reviewer_BzpXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses spatially heterogeneous degradations in all-in-one endoscopic image restoration. The authors identify a potential failure mode termed restoration-action spillover, where corrective operations extend beyond degraded regions and unnecessarily modify intact tissue. To address this issue, the paper proposes BARS-MambaIR, a blind action-regulated state-space restoration framework. A lightweight controller predicts three spatial fields—demand, scope, and magnitude—to regulate state writing, retention, boundary attenuation, multi-range routing, and final residual readout. These fields are learned using pseudo-supervision constructed from paired input–reference residuals, without requiring explicit degradation labels or manually annotated artifact masks. The method is evaluated on SurgClean and SegSTRONG-C against several existing restoration approaches. The experiments demonstrate competitive restoration performance and reduced restoration-action spillover. Additional analyses examine restoration–preservation trade-offs, degradation geometry, controller interventions, and spatial routing behavior.

Critical strengths
  1. Meaningful problem formulation and well-motivated framework. The paper addresses an important issue in all-in-one image restoration: restoring degraded regions while minimizing unnecessary modifications to intact image content. The restoration-action spillover perspective provides a useful motivation for studying spatially adaptive restoration beyond conventional global image-quality metrics. The proposed demand–scope–magnitude framework is conceptually coherent. Its integration with state-space dynamics provides an interesting approach to controlling where, how far, and how strongly restoration operations are applied.
  2. Promising restoration performance and informative analyses.** The proposed method achieves competitive restoration results on two benchmarks, with improvements over several established methods. The additional restoration–preservation analyses, demand-field interventions, and support-dependent routing statistics provide useful insights into the model's spatially adaptive behavior. These results suggest that the overall framework is promising and worthy of further investigation.
Critical weaknesses
  1. Insufficient validation of the reliability and semantic validity of the residual-derived pseudo-supervision. A central concern is the lack of independent validation of the pseudo-supervision used to learn the demand, scope, and magnitude fields. As defined in Eq. (15), the supervision targets are constructed through heuristic transformations of paired input–reference residuals rather than obtained from independent ground-truth annotations or measurements. Specifically, demand and magnitude targets are derived from residual amplitudes, while the scope target is determined by the relative contributions of low- and high-frequency residual components. Although these formulations are intuitively motivated, the paper does not sufficiently demonstrate that the constructed targets reliably correspond to their intended meanings. In particular, the assumption that residual frequency characteristics reflect the appropriate state propagation range requires further justification. The reported restoration improvements demonstrate the practical effectiveness of the complete training framework, but they do not necessarily establish the accuracy or semantic validity of the constructed pseudo-targets. Similarly, the controller interventions and routing statistics indicate that the learned fields are useful and exhibit plausible behavior, but do not directly validate the reliability of the underlying pseudo-supervision. Since these fields are fundamental to the proposed framework and its mechanistic interpretation, independent validation would substantially strengthen the paper. This could include comparisons with independently annotated degradation regions, controlled degradations with known spatial characteristics, or quantitative analyses relating the predicted scope to empirically beneficial propagation ranges.
  2. Insufficient ablation studies to establish the necessity and independent contributions of the proposed components.** The proposed framework combines several mechanisms, including demand-based state gating, scope-dependent state retention, boundary-conditioned attenuation, multi-range routing, and output residual gating. However, the current experiments do not sufficiently isolate the individual contributions of these components. In particular, Eq. (14) directly modulates the final restoration residual using the predicted demand and magnitude fields. This output-level gating mechanism could itself reduce unnecessary modifications in intact regions, independently of the proposed regulation of internal state-space dynamics. Although the demand-field intervention experiments demonstrate the importance of spatially aligned control, they do not distinguish the contributions of internal state regulation and output gating. Furthermore, removing the explicit boundary term appears to have only a limited effect on the reported aggregate metrics, leaving its independent contribution unclear. More systematic ablations are needed to establish the necessity of the major components. Particularly important comparisons include variants without the pseudo-supervision losses, variants using alternative pseudo-target constructions, and output-gating-only versus state-regulation-only variants under comparable training conditions.
Questions for authors
  1. Validation of the pseudo-supervision. How do the authors validate that the residual-derived pseudo-targets in Eq. (15) reliably represent their intended meanings, particularly the relationship between residual frequency characteristics and the appropriate propagation scope? Could the authors provide independent quantitative evidence of the semantic validity of these targets, for example using a small annotated subset, controlled synthetic degradations, or analyses relating the predicted control fields to independently measured restoration requirements?
  2. Ablation of the proposed control mechanisms. Could the authors provide more comprehensive component-level ablation studies to clarify the individual contributions of demand regulation, scope-dependent retention, boundary attenuation, multi-range routing, and magnitude control? In particular, how do output-gating-only, state-regulation-only, and full-model variants compare under similar model capacities and training conditions?

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

Related papers

Discussion (0)

Sign in to comment.