- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes ReDU-IR, an all-in-one blind image restoration framework based on Reconstructible Feature Decomposition (RFD) and State Complementary Updating (SCU). RFD decomposes encoder features into main and complementary representations and reconstructs them through an inverse transform. SCU updates the complementary features based on the differences between encoder and decoder main states. Experiments on multiple restoration benchmarks demonstrate competitive performance, particularly under composite degradations.
Critical strengths
- The paper addresses a meaningful problem of information preservation and feature reuse in hierarchical image restoration networks. The proposed RFD and SCU modules are well motivated and technically sound, particularly the use of paired feature reconstruction and state-dependent complementary updates.
- The experimental results are strong across several restoration benchmarks, particularly CDD11 and WeatherBench. The relatively comprehensive ablation studies further support the effectiveness of the proposed components.
Critical weaknesses
- The methodological novelty of RFD is somewhat limited, as invertible and wavelet-based representations have been explored in previous restoration methods. Moreover, the nearly perfect feature reconstruction is an expected property of invertible transformations, and its practical advantages over conventional feature reuse are not sufficiently demonstrated.
- The effectiveness of SCU and the consistency of the improvements require further justification. In Table 5, SCU provides only a 0.08 dB improvement in average PSNR, raising questions about its additional complexity. Furthermore, the method falls behind several existing approaches on deblurring and low-light enhancement in the 5D setting, suggesting that its benefits may be task-dependent.
Questions for authors
- What are the main advantages of RFD compared with existing wavelet-based or invertible feature decomposition methods, beyond the learnable orthogonal transformation?
- Given the relatively small improvements from SCU in Table 5, could the authors further explain why the state-dependent update is necessary compared with simpler feature fusion strategies?
- Could the authors provide additional analysis showing how the reconstruction property of RFD affects restoration quality, particularly the preservation of high-frequency details?
- Why does ReDU-IR show relatively limited improvements on deblurring and low-light enhancement compared with its stronger results on other degradation types?
3 comments
Response Part 2 of 3 — Q2–Q3
Q2:
Weakness 2: The effectiveness of SCU and the consistency of the improvements require further justification. In Table 5, SCU provides only a 0.08 dB improvement in average PSNR, raising questions about its additional complexity. Furthermore, the method falls behind several existing approaches on deblurring and low-light enhancement in the 5D setting, suggesting that its benefits may be task-dependent.
Question 2: Given the relatively small improvements from SCU in Table 5, could the authors further explain why the state-dependent update is necessary compared with simpler feature fusion strategies?
A2:
Thank you for your suggestion. SCU updates the complementary features using the difference between shared RSR responses to the encoder and decoder main states. This ties the update to the change in the main representation during restoration. We have repeated all five SCU ablations using the same three random seeds and now report mean PSNR ± standard deviation in Tab. 5(b). Compared with concatenation-based fusion, SCU raises mean PSNR from 33.15 ± 0.025 to 33.22 ± 0.036 dB with 0.20M additional parameters. Decoder-only updating achieves 33.16 ± 0.045 dB at the same parameter count as SCU.
Table 5(b). SCU ablation in the ReDU-IR-B configuration under the 3D setting, with the learned RFD fixed. All five variants use the same training protocol and the same three random seeds. PSNR is reported as mean ± standard deviation.
| Variant | Module configuration | Params. (M) | PSNR (dB) ↑ |
|---|---|---|---|
| w/o SCU | Direct complementary feature reuse | 21.95 | 33.13 ± 0.034 |
| w/o encoder-side RSR | Decoder-only RSR | 22.54 | 33.16 ± 0.045 |
| w/o SCU (fusion replacement) | Concatenation-based fusion | 22.34 | 33.15 ± 0.025 |
| w/o RSR weight sharing | Separate encoder and decoder RSRs | 23.14 | 33.19 ± 0.038 |
| ReDU-IR (full model) | SCU with shared RSR | 22.54 | 33.22 ± 0.036 |
We have updated Table 5(b) and the “SCU Update Design” discussion in Section 4.2 (page 9) with three-seed statistics and comparisons with alternative update mechanisms.
Q3:
Question 3: Could the authors provide additional analysis showing how the reconstruction property of RFD affects restoration quality, particularly the preservation of high-frequency details?
A3:
We thank you for this suggestion. We further explain the role of reconstructibility in terms of information preservation and restoration performance. RFD distributes high-frequency information between the main and complementary representations and recombines them at the decoder through a matched inverse transform. This allows information that does not enter the main restoration pathway to still contribute to subsequent reconstruction, thereby preserving information that can support the restoration of fine details. When both the main and complementary representations are left unchanged, the relative reconstruction error for the illustrated sample is , verifying this information-preserving property.
In restoration experiments, Table 15 shows that adding RFD without SCU improves PSNR by 1.43 dB on CDD11 and 1.14 dB on WeatherBench, supporting the contribution of the RFD module to restoration quality. Table 5(a) further shows that, while maintaining a reconstructible decomposition, RFD with learnable high-frequency allocation improves PSNR by 0.11 dB over fixed Haar decomposition. Together, these results support the RFD design in terms of both information preservation and restoration performance: matched reconstruction retains detail information for subsequent restoration, while learnable high-frequency allocation adapts the organization of this information to the restoration objective.
The reconstruction evidence in Section 3.2/Figure 3 (page 5) and Table 5(a) (Section 4.2, page 9) was already in the submission. We have added the component ablation in Table 15 (Appendix F.2, page 19).
Response Part 3 of 3 — Q4
Q4:
Question 4: Why does ReDU-IR show relatively limited improvements on deblurring and low-light enhancement compared with its strong results on other degradation types?
A4:
We thank you for this observation. We agree that ReDU-IR does not achieve the best results on deblurring and low-light enhancement, but it is useful to distinguish improvements over the baseline from rankings against other methods on individual tasks. Compared with SymUNet, ReDU-IR-B improves deblurring PSNR from 28.12 to 28.75 dB (+0.63 dB) and SSIM from 0.855 to 0.871 (+0.016), showing clear improvements in both metrics for this task.
The five-task results also show that different unified restoration methods have different task-specific strengths. For example, MoCE-IR achieves the highest deblurring PSNR, whereas ReDU-IR-B achieves higher PSNR on each of the other four tasks. MIRAGE-S and DFPIR achieve higher low-light PSNR, but perform worse than ReDU-IR-B on several other tasks. ReDU-IR-B achieves the highest average PSNR across the five tasks, at 31.03 dB. These results indicate strong overall restoration performance, although the gains are uneven across tasks.
On low-light enhancement, PSNR increases by 0.20 dB over SymUNet, while SSIM decreases by 0.001. These comparisons establish uneven task-specific gains but do not isolate their causes.
5D comparison from revised Table 2. Entries are PSNR (dB) / SSIM; bold indicates the better score between the two models. Denoising is evaluated on CBSD68 (σ = 25).
| Model | Params. (M) | Dehazing | Deraining | Denoising | Deblurring | Low-Light | Average |
|---|---|---|---|---|---|---|---|
| SymUNet | 22M | 31.31/0.979 | 38.05/0.981 | 31.38/0.891 | 28.12/0.855 | 23.27/0.858 | 30.43/0.913 |
| ReDU-IR-B (Ours) | 23M | 32.29/0.982 | 39.10/0.985 | 31.56/0.894 | 28.75/0.871 | 23.47/0.857 | 31.03/0.918 |
We have added SymUNet (22M), its per-task and average PSNR/SSIM scores, and the Du et al. (2026) citation to Tables 1 and 2 in Section 4.1 (pages 7–8).
Response Part 1 of 3 — Q1
Thank you for your valuable comments, which have helped us improve our manuscript. We have revised the manuscript based on your feedback. Below, we provide details of our revisions. Please let us know whether we have addressed your concerns.
Q1:
Weakness 1: The methodological novelty of RFD is somewhat limited, as invertible and wavelet-based representations have been explored in previous restoration methods. Moreover, the nearly perfect feature reconstruction is an expected property of invertible transformations, and its practical advantages over conventional feature reuse are not sufficiently demonstrated.
Question 1: What are the main advantages of RFD compared with existing wavelet-based or invertible feature decomposition methods, beyond the learnable orthogonal transformation?
A1:
We thank you for this helpful comment. We agree that near-perfect reconstruction of unchanged features is an expected property of invertible transforms. The contribution of RFD lies in how it organizes information for restoration while retaining a matched reconstruction path.
MWCNN (Liu et al., 2018) uses DWT/IWT for resolution changes, while WINNet (Huang & Dragotti, 2022) combines a lifting-inspired transform (LINN) with detail-coefficient denoising. In RFD, a portion of the mixed high-frequency coefficients is processed jointly with the low-frequency component along the main restoration pathway; the remaining coefficients form a complementary representation for matched reconstruction. This design specifies which information receives deeper restoration and which information is retained outside that pathway. The learnable orthogonal transform adapts the high-frequency combinations assigned to each representation while retaining the matched inverse transform. In the full ReDU-IR model, SCU additionally updates the complementary representation before synthesis; its contribution is evaluated separately from RFD.
Without SCU, Table 5(a) shows that RFD improves PSNR over the SymUNet PixelUnshuffle baseline from 32.95 to 33.17 dB (+0.22 dB), with 0.34M fewer parameters. RFD also exceeds fixed Haar decomposition by 0.11 dB. Table 15 further shows RFD-only PSNR gains of 1.43 dB on CDD11 and 1.14 dB on WeatherBench. These results support the practical value of the RFD design in restoration.
We additionally compare RFD with DWT/IWT and LINN as replacement modules within SymUNet (dim=32), without SCU (Table 17 below). RFD achieves 31.62 dB/0.881 with the fewest parameters (9.87M) and FLOPs (35.07G). Its quality gain over LINN is modest (+0.06 dB/+0.001), accompanied by lower latency (17.11 versus 39.97 ms). DWT/IWT is faster (14.47 ms), while RFD improves PSNR/SSIM by 1.33 dB/0.015. This comparison supports a favorable balance between restoration quality and computational cost in the evaluated backbone.
Table 17. RFD versus DWT/IWT and LINN on WeatherBench (SymUNet, dim=32; without SCU).
PSNR/SSIM are averaged over haze, rain, and snow. Params., FLOPs, and time refer to complete models. Time is latency in ms on an RTX 4090 with 256×256 inputs and batch size 1. FLOPs are reported using the THOP MAC-count convention.
We have clarified RFD’s design in the final paragraph of Section 2 (page 3) and added the component and wavelet-module comparisons in Appendix F.2 (Tables 15 and 17, page 19; analysis, pages 19–20).