acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Wavelet Attention-Guided Tensor Completion with Fused Low-Rankness and Smoothness for Hyperspectral Image Restoration

Abstract

Low-rank tensor completion is widely used to recover missing data in hyperspectral images. However, existing methods often rely on fixed low-rank models and hand-crafted priors, which may fail to preserve spatial and spectral structures when large amounts of data are missing. To address this issue, we propose Adaptive Wavelet Attention-guided Smooth Tensor Completion (AWSTC), an adaptive wavelet attention-guided deep unfolding framework for recovering missing entries in hyperspectral image tensors by jointly exploiting low-rankness and smoothness. AWSTC models tensor completion as a low-rank factorization problem with fused smoothness regularization. Its adaptive wavelet attention module learns the importance of different channels from wavelet subbands to guide low-rank recovery. The resulting optimization algorithm is unfolded into a multi-stage network that combines model-based updates with learnable residual dense blocks. Experiments on three hyperspectral image datasets show that AWSTC performs well across various missing rates, producing clearer images with fewer stripe artifacts.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hyperspectral image restoration, Low-rank tensor completion, Deep unfolding, Adaptive wavelet attention, Fused low-rankness and smoothness

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uX6vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes AWSTC for hyperspectral image completion by combining gradient-domain low-rank tensor factorization, learned auxiliary mappings, and adaptive wavelet attention. Haar-initialized learnable analysis and synthesis filters, high-frequency thresholding, and subband attention generate channel-wise weights for the low-rank fitting term. The proposed variable updates are unfolded into a multi-stage network. Experiments on Indian Pines, Salinas, and Pavia University report strong recovery at several missing rates, with comparisons of attention modules, individual AWA components, and unfolding depth.

Critical strengths
  1. The motivation for combining gradient-domain low-rank structure with frequency-aware attention is clear, and the AWA construction is specified beyond a generic attention block. The manuscript explicitly acknowledges that its learned filters need not be orthogonal or perfectly reconstructive, and that the CNN mappings are not claimed to be exact proximal operators of explicit regularizers.

  2. The submission reports strong restoration results and includes meaningful attention controls. Table 4 compares several attention alternatives with the remaining components fixed, while Table 5 separately examines learned filters, thresholding, subband attention, and high-frequency fusion. These are useful positive observations, although their interpretation remains subject to the evaluation concern below.

Critical weaknesses
  1. The factor updates do not follow from the stated weighted objective. Equations (14)–(15) weight residuals on the original spectral channels, but (B.6) removes those weights before deriving the unweighted Fourier-slice updates in (17)/(B.8)–(B.10). Nonuniform original-band weights generally produce a nondiagonal metric after the mode-3 DFT and couple Fourier slices; positivity alone does not remove them. A small rank-one, two-band example gives a fixed point of the stated unweighted updates with a nonzero gradient of the original weighted objective. Thus, the claimed derivation of these model-based stages is not established. This does not imply that AWA is absent from the network—(21) retains it—or that the learned architecture cannot work as a heuristic.

  2. The supervised training and evaluation split is not established, and the stated Indian Pines dimensions make a disjoint same-scene crop protocol impossible. Section 4.1 gives a 145×145 original image, a 128×128 evaluation crop, and 32×32 training patches sampled from the original HSI with ground-truth supervision. Any such axis-aligned patch must overlap the evaluation crop by at least 15 pixels in each dimension. The submission does not specify another training source or an exclusion policy. This does not confirm what the actual implementation did, but the written protocol cannot support an interpretation as evaluation on spatially unseen content. Model/depth selection and learned-baseline training access also require clarification.

Questions for authors
  1. How do (17)/(B.6)–(B.10) follow from the channel-weighted term in (14)? Please specify the domain of the weights and the implemented factor update. If the unweighted step is an intentional approximation, provide its relation to the objective and revise the derivation claim; if the forward update must change, please identify and update the affected results.

  2. Please provide the training/validation/test sources and crop coordinates or exclusion rules for each dataset, especially explaining the 145/128/32 Indian Pines dimensions. Were clean values from the evaluated region used for supervised fitting? If this is intentional same-scene adaptation, please define that setting, establish comparable clean-data access for the baselines, and provide an evaluation consistent with the claimed task. Please also identify the independent data, if any, used to choose checkpoints and unfolding depth.

3 comments

We thank the reviewer for the careful assessment and for recognizing our motivation and attention ablations. We clarify the factor updates, revised spatial partition, and model-selection protocol below.

1. Factor updates and the weighted objective

Our factor-update construction follows Mai et al., “Attention-Guided Low-Rank Tensor Completion” (AGTC), which formulates an attention-weighted factor subproblem in Eq. (16) and provides Fourier-slice pseudoinverse updates in Eqs. (17)–(18). We adapted this construction to directional-gradient tensor factorization.

In AWSTC, AWA weights are defined on the original spectral channels and broadcast spatially. The factor step performs unweighted least-squares fitting, while the gradient-tensor update explicitly incorporates AWA weights to adjust channel-wise fitting contributions. Together, these steps constitute the unfolding stage.

We recognize that the reviewer's question concerns exact minimization of the weighted subproblem, rather than the presence of attention. We therefore distinguish the construction's provenance from its optimization interpretation: the adopted factor step is an unweighted surrogate update, not an exact solution of the generally nonuniform channel-weighted subproblem. The AGTC reference establishes the source of this construction; it does not itself prove equivalence for arbitrary channel weights.

2. Revised spatially disjoint training/test partition

Thank you for identifying the spatial-overlap issue. The original sampling procedure allowed overlap between training patches and the evaluation region. We have corrected the partition and rerun the experiments using spatially non-overlapping training and test regions.

The revised protocol is specified in Section 4.1:

  • Indian Pines: The upper-left 64 × 128 × 200 subcube is reserved for testing. Training patches of size 32 × 32 × 200 are sampled only from rows 65–145.
  • Salinas: The upper-left 192 × 192 × 224 subcube is reserved for testing. Training patches of size 64 × 64 × 224 are sampled only from rows 218–512 across the full image width.
  • Pavia University: The upper-left 256 × 256 × 103 subcube is reserved for testing. Training patches of size 64 × 64 × 103 are sampled only from rows 257–609 and columns 257–339.

All spectral bands are retained. For Indian Pines, replacing the original 128 × 128 test crop with a 64 × 128 crop leaves a separate lower region accommodating 32 × 32 training patches, resolving the dimensional conflict identified by the reviewer.

Under this protocol, clean reference values from the test region are excluded from supervised parameter updates. The experiment evaluates restoration in spatially held-out regions within each scene, rather than cross-scene generalization. The revised partition is highlighted in blue in Section 4.1, and updated results appear in Tables 1–3 and the subsequent experimental analyses.

3. Checkpoint selection, unfolding depth, and baseline training data

The final checkpoint was selected as the training epoch with the highest PSNR on the designated test region. The depth of 11 stages was also selected according to PSNR on that region. No separate validation region was used for either choice.

Thus, although the test region was excluded from supervised parameter updates, its clean reference values were used for checkpoint and depth selection. The corrected spatial split removes direct overlap between training patches and test pixels, but does not eliminate this model-selection dependence.

AWSTC, AGTC, and the other learning-based methods in this comparison use the same training data, providing a common supervised training-data basis. This condition does not, by itself, imply identical checkpoint or hyperparameter selection procedures.

We appreciate the opportunity to clarify these distinctions and the scope of the reported results.

Reviewer_uX6vOpenReview ↗

Reply to Authors' Response

I appreciate the authors' efforts in providing a detailed reply and experiments. However, my concern still remains, especially in the independent evaluation: In my point of view: the reported scores somehow do not provide an independent estimate of performance after model selection. Specifically, the revised split addresses spatial overlap, but the authors confirm that both the checkpoint and the 11-stage depth were selected using ground-truth PSNR on the same region used to report test performance. I might incline to my current rate if these ambiguities cannot be resolved.

Thank you for clarifying this concern. We agree that spatial separation alone does not provide an independent performance estimate when the same region is used for checkpoint and depth selection. We have therefore added an evaluation on a previously unused region, reported in Section 4.6 and Table 7.

Independent evaluation protocol. On Pavia University at 50% missingness, we evaluate region B: rows 1–192 and columns 257–320 (one-based indexing), with size 192 × 64 × 103. This region does not overlap either the training region or the original model-selection region. It was not used for training, checkpoint selection, depth selection, or other parameter tuning.

Before evaluating region B, we fixed the previously selected 11-stage AWSTC and its Pavia University checkpoint, without retraining or further model selection. All ten methods were evaluated on the same region, corrupted input, and missing-data mask, using settings fixed before evaluation. No parameters or outputs were selected using region B’s ground truth. We reconstructed three non-overlapping 64 × 64 blocks and computed the metrics after concatenating them into the full region.

Results. AWSTC achieves 48.11 dB PSNR, 0.99 SSIM, 0.99 FSIM, and 21.43 ERGAS. Compared with CRNL, the strongest baseline in PSNR and ERGAS, AWSTC improves PSNR by 3.76 dB and reduces ERGAS by 25.2%. SSIM and FSIM are joint-best at the reported precision.

This experiment evaluates the already selected model on data excluded from both training and model selection, directly addressing the concern about an independent performance estimate. We explicitly distinguish the original region, which served model selection, from the new independent test region. The evidence is limited to this additional within-scene evaluation on Pavia University at 50% missingness; it does not make the earlier results independent of model selection or establish cross-scene generalization.

The protocol and results are highlighted in blue in the revised manuscript. We hope this additional evidence addresses your concern and would appreciate your reconsideration of the assessment.

Reviewer_bd6MOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the AWSTC framework for hyperspectral image completion and stripe noise removal. The framework integrates directional gradient-domain low-rank matrix factorization with bounded noise relaxation, introduces a learnable Adaptive Wavelet Attention module to adaptively extract spectral band weights, and employs ADMM to unroll the algorithm into an end-to-end deep neural network incorporating Residual Dense Blocks.

Critical strengths
  1. Introduces the Adaptive Wavelet Attention module into the low-rank tensor completion model, thereby guiding spectral band weight allocation.

  2. Completely maps the constrained optimization problem into an end-to-end trainable deep unrolling network based on the ADMM algorithm, replacing proximal operators with RDBs.

Critical weaknesses
  1. Training patches are randomly cropped directly from the same hyperspectral image used for testing, leading to severe data leakage. The network memorizes test scene structures and yields unrealistically high PSNR values.

  2. In the ablation study, increasing the number of unrolling stages from 9 to 11 causes an abrupt PSNR jump of nearly 10 dB, followed by a rapid drop at stage 12, which violates the smooth convergence behavior of unrolled networks and indicates severe overfitting or instability.

  3. The observation constraint in Equation (14) contains a logical flaw: the mask operator on the right forces missing entries to zero, while the reconstructed tensor on the left is dense.

  4. The paper lacks quantitative comparisons with recent deep learning SOTA baselines (2024–2026), as almost all benchmarked algorithms are traditional unsupervised tensor factorization models.

Questions for authors
  1. Could the authors clarify whether spatial overlap exists between training patches and test regions, and provide results under a strict disjoint scene split?

  2. What causes the drastic PSNR jump of nearly 10 dB when increasing the stage count from 9 to 11?

2 comments

Response to Concern 1: Training/test overlap

Thank you for identifying this important issue. We acknowledge that the original patch-sampling procedure allowed spatial overlap between training patches and the test region. We have corrected the partitioning and rerun the experiments using spatially non-overlapping training and test regions.

The revised protocol is detailed in Section 4.1 (Setup):

  • Indian Pines: The upper-left 64 × 128 × 200 subcube is reserved for testing. Training patches of size 32 × 32 × 200 are sampled only from rows 65–145.
  • Salinas: The upper-left 192 × 192 × 224 subcube is reserved for testing. Training patches of size 64 × 64 × 224 are sampled only from rows 218–512 across the full image width.
  • Pavia University: The upper-left 256 × 256 × 103 subcube is reserved for testing. Training patches of size 64 × 64 × 103 are sampled only from rows 257–609 and columns 257–339.

No training patch contains pixels from the test regions. This is a spatially disjoint split within each scene, rather than a cross-scene evaluation. The revised protocol is highlighted in blue in Section 4.1, and updated results are reported in Tables 1–3 and the subsequent experimental analyses. We thank the reviewer for prompting this correction.

Response to Concern 2: Depth ablation and overfitting

Thank you for raising this concern. The depth ablation compares separately trained networks with different unfolding depths, rather than intermediate outputs of a single network. Therefore, it does not directly establish or contradict within-network convergence.

To investigate overfitting, we evaluated fixed checkpoints on the same 100 training patches and the same spatially held-out region, without updating parameters. Across depths, evaluation inputs, missing-data masks, and PSNR calculations were consistent. Results are reported in Section 4.5 and Table 6.

From 9 to 11 stages, training PSNR increased from 43.2396 to 52.2883 dB, while held-out PSNR increased from 42.1449 to 52.1423 dB. The training–held-out gap decreased from 1.0947 to 0.1460 dB. Thus, the improvement also occurred on data excluded from training, providing evidence against pronounced depth-related overfitting in this setting.

From 11 to 12 stages, held-out PSNR decreased by only 0.0668 dB, from 52.1423 to 52.0755 dB, while training PSNR changed marginally. This suggests performance saturation around 11 stages rather than a sharp deterioration.

Response to Concern 3: Observation constraint

Thank you for pointing out this issue. We agree that unobserved entries of the reconstructed tensor should not be forced to zero. Section 3.2 now explicitly clarifies that the slack tensor is bounded only on observed entries.

The revised formulation requires X + E + S to equal the zero-filled observations, while bounding only the observed component of S. Eliminating S yields an equivalent constraint that bounds the reconstruction residual only at observed entries.

At unobserved entries, S equals the negative of X + E. Since this component of S is unrestricted, it absorbs the corresponding values without forcing the reconstruction to zero. The zero-filled right-hand side therefore does not impose zero values on missing entries.

Consistently, the slack update preserves its unobserved component and projects only its observed component onto the prescribed Frobenius-norm ball. The clarification is highlighted in blue in Section 3.2; detailed derivations are in Appendix B.

Response to Concern 4: Recent learning-based baselines

Thank you for this suggestion. We have added CRNL (2025) and 3DeepRep (2026) to broaden the comparison with recent learning-based methods.

The baseline list is updated in Section 4.1. Tables 1–3 report PSNR, SSIM, FSIM, and ERGAS on all three datasets at 50%, 70%, and 90% missingness under the same degradation settings. The added methods also appear in Figure 3 and the error maps in Appendix C, Figure 6.

AWSTC achieves the highest PSNR in eight of nine settings. On Salinas at 70% missingness, AWSTC achieves 51.25 dB, compared with 51.07 dB for CRNL and 49.48 dB for 3DeepRep.

The added baseline entries, quantitative results, and corresponding discussion are highlighted in blue.

Reviewer_bd6MOpenReview ↗

Remaining Concerns

Thank you for the revisions. The corrected spatial split, constraint clarification, and additional baselines address several concerns. However, the main benchmark results remain affected by using test-region ground truth for checkpoint and depth selection. The new independent evaluation is helpful, but covers only Pavia University at 50% missingness and does not resolve this issue for the broader benchmark. I therefore maintain my rating.

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

This manuscript presentsa deep unfolding framework for hyperspectral tensor completion that brings together low-rank factorisation, smoothness regularisation and wavelet-based attention.

Critical strengths

Combining model-driven optimisation with learned representations is a potentially useful approach.

Critical weaknesses

The logic of contribution looks questionable: the author said “AWSTC models tensor completion as a low-rank factorization problem with fused smoothness regularization”, while we already have lots of this kinks of framework in the literature, for example varied low-rankness terms with tv or tgv to model smoothness. If we look at the wavelet attention + low rankness framework that used in this paper, we will find that the framework appears closely related to existing methods that combine low-rank tensor factorisation, attention mechanisms and deep unfolding, in particular Attention-Guided Low-Rank Tensor Completion (TPAMI 2024).

Second, the author said the GAP of existing methods is that “existing methods often rely on fixed low-rank models and hand-crafted priors”, I don’t think this is true, given that PnP and deep unfolding methods already been proposed for almost 10 years.

Third, wavelet based attention also has been widely used in general image restoration tasks, e.g., the SR, denoising\ debluring in low-level vision.

Finally, some minor comments: the text within figure-1 is too small, figure 2 has the same issue, the reference of the introduction is too old, most of them are published by 2020, some equations are with comma in the end, but some equations are without.

Questions for authors

See Weaknesses part.

2 comments

We thank the reviewer for the careful assessment and for recognizing the value of combining model-driven optimization with learned representations. Your comments have helped us clarify our contributions and improve the discussion of related work. We respond below.

1. Contribution and relationship to AGTC

We agree that combining low-rank modeling with smoothness regularization, or attention with deep unfolding, is well established. Our contribution lies in the specific integration of directional-gradient low-rank modeling and adaptive wavelet-based channel weighting for hyperspectral image completion.

AGTC is an important related work and the source of our factor-update construction. The main distinctions are:

  • Object of low-rank modeling: AGTC applies low-rank factor fitting to the reconstructed tensor. AWSTC instead factorizes directional-gradient tensors and links them to the reconstruction through gradient-consistency constraints, jointly modeling gradient correlations and spatial–spectral smoothness.
  • Attention construction and role: AWA combines learnable wavelet-like analysis, subband thresholding, directional subband attention, and high-frequency fusion to generate spectral-channel weights. These weights modulate gradient-domain low-rank fitting and are shared across unfolding stages.

We therefore position AWSTC as a specific extension of attention-guided tensor restoration. The gradient-domain formulation and AWA construction are detailed in Sections 3 and 3.1, respectively.

Table 4 compares attention modules with the remaining components fixed. On Indian Pines at 50% missingness, no attention, AHDRNet attention, CBAM, SimAM, and AWA achieve PSNR values of 46.11, 46.78, 48.12, 51.14, and 52.14 dB, respectively. These results support the effectiveness of AWA within our architecture.

2. Characterization of existing methods

We accept this criticism. Broadly describing existing methods as relying on fixed low-rank models and handcrafted priors does not adequately reflect developments in PnP, deep unfolding, and learning-based tensor restoration.

The revised Introduction explicitly discusses learned priors, deep unfolding, and attention-guided tensor restoration, with additional recent references. Our motivation is more precisely to investigate how directional-gradient modeling and frequency-aware channel weighting can adapt restoration to different hyperspectral structures within an established model-and-learning framework. The relevant revisions appear in Section 1.

3. Relationship to existing wavelet-attention methods

We agree that wavelets and attention have been widely used in super-resolution, denoising, and deblurring. Using wavelet attention alone is therefore not our novelty claim.

Our proposed AWA module contributes a specific construction and its integration into gradient-domain tensor restoration. Its analysis and synthesis filters are initialized with Haar bases and subsequently learned. Subband-adaptive thresholding, directional subband channel attention, and high-frequency fusion jointly generate spectral-channel weights for gradient-domain fitting.

Table 5 examines these components individually. Under the same setting, complete AWA achieves 52.14 dB. Fixing the filters to Haar, removing subband attention, removing high-frequency fusion, and removing thresholding yield 51.44, 50.91, 51.99, and 51.89 dB, respectively. These ablations provide empirical support for the specific design. Details appear in Sections 3.1 and 4.4 and Tables 4–5.

4. Figures, recent references, and equation punctuation

Thank you for these concrete suggestions.

  • Figure readability: We have updated Figures 1 and 2 and displayed them at full text width to improve the readability of the structural illustrations and framework details.
  • Recent literature: Section 1 now includes additional studies from 2024–2026 covering low-rank gradient models, spatial–spectral deep priors, 3DeepRep, and hyperspectral deep unfolding. We have also added quantitative and visual comparisons with CRNL (2025) and 3DeepRep (2026), reported in Tables 1–3 and Figure 3.
  • Equation punctuation: We have revised punctuation according to each equation's grammatical role, using commas within sentences and periods at sentence endings.

We appreciate the reviewer's suggestions. The revision more precisely identifies our contribution in directional-gradient modeling, the proposed AWA construction, and its empirical effects, while clarifying the relationship to prior work.

Reviewer_kcTjOpenReview ↗

Novelty

Thank you for the rebuttal and the additional clarifications. I appreciate the authors' efforts to address my concerns. However, I remain unconvinced by the methodological novelty. As the authors acknowledge, combining low-rank modeling with smoothness regularization, as well as attention mechanisms with deep unfolding, is already well established. The proposed integration of directional-gradient low-rank modeling and adaptive wavelet-based channel weighting appears to be an incremental extension of existing approaches rather than a fundamentally new contribution. While the clarification is helpful, it does not sufficiently address my original concern. I therefore prefer to maintain my initial score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.