acceptodds
Under review as a conference paper at ICLR 2027

QuHIC: Quantization-Consistent Multi-Prior Unrolled Reconstruction for Hyperspectral Image Compression

Abstract

Recent advances have demonstrated the great potential of lossy hyperspectral image (HSI) compression for airborne remote-sensing applications. However, given the limited storage capacity and transmission bandwidth of airborne sensing platforms, most existing methods insufficiently integrate compression-induced constraints with spectral–spatial priors, making it difficult to balance low bit rates and high-fidelity reconstruction. To address this challenge, we propose QuHIC, which synergizes bit-level quantized compression with multi-stage unrolled reconstruction, establishing a new paradigm for HSI compression and reconstruction. Specifically, during encoding, QuHIC employs a learnable lightweight encoder to compress an HSI into a low-dimensional quantized representation for efficient storage and transmission. During decoding, HSI reconstruction is formulated as an iterative optimization problem jointly constrained by quantization consistency and multi-prior regularization. Each iteration alternates between a quantization-consistency module (QCM) that explicitly corrects deviations from the feasible quantization interval and a multi-prior refinement network (MPRN) that serves as a learnable proximal operator integrating raw-band spectral continuity and curvature modeling, multi-scale spatial detail enhancement, and gated spectral–spatial contextual interaction. Extensive experiments demonstrate that the proposed QuHIC not only achieves efficient HSI compression at competitive bit rates but also outperforms many state-of-the-art methods in both reconstruction fidelity and spectral detail preservation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hyperspectral Image Compression, Hyperspectral Image Reconstruction, Deep Unrolling.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CptxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses low-bit-rate lossy hyperspectral image compression. A lightweight convolutional encoder with a sigmoid activation maps the input into a compact latent that is uniformly quantized to b bits. On the decoder side, reconstruction is formulated as an optimization problem with a quantization-consistency fidelity term and multi-prior regularization, and is unrolled into nine stages. Each stage alternates a quantization-consistency module, which re-encodes the current estimate, projects it onto the feasible quantization interval, and back-projects only the interval violation, with a multi-prior refinement network. Experiments on AVIRIS, WHU-Hi, and geographically unseen AVIRIS regions evaluate spatial and spectral reconstruction fidelity.

Critical strengths
  1. Treating each quantized value as a feasible interval rather than a single target, and correcting only interval violations inside an unrolled decoder, is a principled way to exploit compression constraints during reconstruction, while keeping the onboard encoder very light. This is the paper's key idea, and it fits the airborne setting well.

  2. The empirical gains are consistent across settings. QuHIC outperforms the strongest baseline, BTC-Net V2, on all five AVIRIS scenes at 32 bits; at 8 bits it improves the average PSNR by 1.04 dB with lower SAM; it remains better at every bit depth from 8 down to 4 bits, with the margin growing to 1.81 dB at 4 bits; and it also leads on WHU-Hi and the unseen regions.

Critical weaknesses
  1. The independent contribution of QCM is not isolated. In Table 3, A0 is a back-projection-only baseline, and A1 adds the multi-stage unrolled decoder and QCM together, so the gain from decoder capacity cannot be separated from the gain from the interval constraint. Since QCM is the main conceptual contribution, an ablation on the full model is needed to support the central claim.

  2. The comparison conditions are not fully matched. The decoder costs 876.53 GFLOPs, about 1.55 times BTC-Net V2 and 7.2 times CTCSN, while the margin over BTC-Net V2 is about 1 dB, so part of the gain may come from additional computation. In addition, bit rates are computed from the latent bit width rather than from the actual Huffman bitstream, and the cross-region and WHU-Hi results are reported only with 32-bit latents, so they do not directly test the low-bit-rate claim.

Questions for authors
  1. On the full model, how does performance compare among (a) removing QCM, (b) back-projecting the residual to the codeword center, and (c) the proposed interval residual?

  2. Could you report a comparison at matched decoding computation, or an ablation on the number of unrolling stages?

  3. What are the actual Huffman bitstream sizes, including side information, for QuHIC and the baselines? Are cross-region results available at low bit depths?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Cptx (1/3)

Thank you for your constructive review and for recognizing the motivation of interval-aware decoding and the reconstruction improvements. We have addressed all concerns below. Revisions in the manuscript addressing the questions below are highlighted in blue. For reference, the original submission with a nine-page main text is also provided in the supplementary material.

R1: QCM ablation and the contribution of interval-based correction

W1 & Q1: With the encoder, MPRN, stages, and training setup fixed, how do no QCM, center-point residual correction, and interval residual correction compare? The A0→A1 result does not separate them.

Response: We appreciate your constructive feedback and apologize for not addressing this point adequately in the original manuscript. We have conducted two sets of experiments: a comparison of three QCM settings and an additional ablation of the basic unrolling baseline. The details are provided below:

1.Comparison of three QCM settings: We conduct the comparison on the AVIRIS test set under 8-bit quantization, keeping the encoder and MPRN architectures, nine-stage unrolling structure, training data, and optimization settings unchanged. Only the QCM operation is varied:

  • No QCM: The QCM residual correction is disabled, while initialization, the MPRN at each stage, and stage fusion are retained.
  • Center-point QCM: The correction target is the quantization center y~\tilde{y}, using the center-point residual fpred−y~f_{\mathrm{pred}}-\tilde{y}.
  • Interval QCM (full QuHIC): The interval residual of the original model is used. The center-point and interval variants employ the same back-projection structure and step-size learning scheme.

Tab. 1. Comparison of three QCM settings under otherwise identical configurations.

IDQCM settingPSNR (dB) ↑RMSE ↓SAM (°) ↓
Q0No QCM39.7654.041.846
Q1Center-point QCM41.4345.981.525
Q2Interval QCM (full QuHIC)41.9743.981.423

Interval QCM improves PSNR by approximately 2.21 dB over no QCM and 0.54 dB over center-point QCM. The former demonstrates the contribution of residual correction within the complete multi-stage framework, while the latter further supports the benefit of using a quantization interval rather than a single center-point target. The rationale is that a quantization symbol specifies a feasible interval, rather than the exact value before quantization.

2.Basic unrolling baseline between A0 and A1: We introduce A0.5: a basic unrolling network without QCM between A0 and A1. This variant adds a basic multi-stage unrolling structure to A0, with all other configurations matched to the original A1.

Tab. 2. Ablation with the additional basic unrolling baseline A0.5.

IDConfigurationPSNR (dB) ↑RMSE ↓SAM (°) ↓
A0Basic compression–backprojection baseline29.77181.835.018
A0.5Basic unrolling network without QCM36.1480.052.607
A1Basic unrolling network + interval QCM37.5567.712.225

Introducing the unrolling structure increases PSNR to 36.14 dB, and adding QCM further improves it to 37.55 dB. This intermediate baseline A0.5 clarifies the sources of the overall A0→A1 improvement.

Manuscript revisions: We have updated Table 3 to include A0.5 and the three QCM comparisons Q0–Q2. We have also moved the ablation visualization to the main text and added ablations of the three QCM settings, as shown in Figure 7. The corresponding analysis has been revised in Sec. 4.3 on p. 9 (Lines 477–481) and p. 10 (Lines 525–531), and in Appendix D.7 on p. 19 (Line 1024), p. 20 (Line 1079), and p. 21 (Lines 1088–1090).

AuthorsOpenReview ↗

Response to Reviewer Cptx (2/3)

R2: Computational trade-off

W2 & Q2: Could you report a comparison at matched decoding computation, or an ablation on the number of unrolling stages?

Response: We appreciate your careful observation and helpful feedback. To clarify the trade-off, we have evaluated additional five- and seven-stage variants and separately profiled the encoder and decoder.

1.Reconstruction quality vs unrolling depth: We evaluate the five-, seven-, and nine-stage variants under the same 32-bit setting, keeping the encoder architecture and latent representation size unchanged.

Tab. 3. Complexity and reconstruction quality under the 32-bit setting.

ModelStagesParams (M)FLOPs (G)PSNR (dB) ↑RMSE ↓SAM (°) ↓
BTC-Net V2-12.550565.0341.3948.361.574
QuHIC56.712495.0541.8246.011.462
QuHIC79.234692.4042.1145.031.416
QuHIC911.580876.5342.5942.881.353

Compared with 32-bit BTC-Net V2, five-stage QuHIC uses 46.5% fewer parameters and 12.4% fewer FLOPs, while improving PSNR by 0.43 dB and achieving lower RMSE and SAM. The seven-stage variant reaches 42.11 dB with an intermediate computational cost, while the full nine-stage model achieves the highest reconstruction quality. These results provide alternatives for different reconstruction budgets.

2.Separate encoder and decoder costs: The encoder has the same complexity across all stage configurations: 0.204 M parameters and 1.674 G FLOPs. The decoder is provided below:

Tab. 4. Separate encoder and decoder network complexity.

StagesEncoder Params (M)Decoder Params (M)Encoder FLOPs (G)Decoder FLOPs (G)
50.2046.5081.674493.375
70.2049.0301.674690.724
90.20411.3761.674874.853

Increasing the unrolling depth adds computation to the reconstruction decoder while leaving the encoding cost unchanged. This reflects our intended deployment design: lightweight neural encoding at the acquisition side and more intensive reconstruction at the ground side.

Manuscript revisions: We have added the five- and seven-stage 32-bit results to Table 1, with the corresponding discussion on p. 7 (Line 377) and p. 8 (Line 418). Their reconstruction visualizations are included in Figure 10, and the encoder–decoder complexity breakdown and corresponding analysis are provided in Table 7 and Appendix D.9 on p. 21 (Lines 1099–1104).

AuthorsOpenReview ↗

Response to Reviewer Cptx (3/3)

R3: Actual Huffman Bitstream

W2 & Q3: What are the actual Huffman bitstream sizes, including side information, for QuHIC and the baselines?

Response: We sincerely apologize for the unclear distinction between fixed-length and actual bitstream bitrates. We clarify the details below:

1.Full-bitstream bitrate: For QuHIC, we measure the complete bitstream size from the actual encoded files, including the Huffman-coded payload, side information required for decoding, and byte-alignment padding. For the HSI with dimensions C×H×WC\times H\times W, let BB denote the total size of the complete bitstream in bytes. The resulting bitrate is R=8BCHWR=\frac{8B}{CHW} bpppb. Huffman coding losslessly recovers the quantized symbols and therefore does not alter the quantized representation supplied to the reconstruction network or the corresponding reconstruction metrics. The updated QuHIC results for the 8-, 7-, and 6-bit settings are listed below, together with the available 8-bit result of BTC-Net V2, which also supports low-bit quantization. The remaining baselines use 32-bit floating-point representations with an uncoded representation bitrate of 0.314 bpppb:

Tab. 5. Bitstream sizes and bitrates.

MethodQuantization bit depthComplete bitstream size (bytes)Bitrate (bpppb)
QuHIC8 bit918030.065
QuHIC7 bit795940.056
QuHIC6 bit692420.049
BTC-Net V28 bit901930.064

2.Role of the 32-bit setting and comparison protocol: The 32-bit setting retains the floating-point latent representation to compare reconstruction capabilities with other floating-point baselines without introducing low-bit uniform quantization. We therefore did not further evaluate the Huffman coding gains for floating-point byte streams. For the low-bit settings, we report the actual bitrate of the complete Huffman-coded bitstream. Thus, the 32-bit results characterize reconstruction performance with a high-precision representation, while the low-bit results evaluate practical compression efficiency.

3.Bitrate Comparison: We have additionally included rate–distortion curves comparing QuHIC with BTC-Net V2, which also supports low-bit quantization. As shown in Figure 5 of the revised manuscript, the curves indicate that QuHIC generally achieves higher PSNR and lower RMSE and SAM at the same bitrate, supporting its better rate–distortion performance.

Manuscript revisions: We have updated the bitrates and comparison statements in Table 1 and in Sec. 4.1 on p. 7 (Lines 373–377). We have added the bitrate computation in Appendix D.4 on p. 17 (Lines 916–917) and p. 18 (Lines 930–949), with the complete-bitstream bitrate formula in Eq. 22 on p. 18 (Lines 945–946). The rate–distortion curves have been added in Figure 5, with the corresponding discussion in Sec. 4.2 on p.8 (Line 481), and p. 9 (Lines 468–470).

R4: Low-bit evaluation on WHU-Hi and unseen AVIRIS

W2 & Q3: Are cross-region results available at low bit depths?

Response: We appreciate this helpful suggestion. We have extended the WHU-Hi and cross-region evaluations to 8-bit quantization, complementing the original 32-bit results.

Tab. 6. Comparison of 32-bit and 8-bit QuHIC.

DatasetRepresentationPSNR (dB) ↑RMSE ↓SAM (°) ↓
WHU-Hi32 bit31.810.2264.466
WHU-Hi8 bit31.690.2234.303
Unseen AVIRIS32 bit41.2123.360.955
Unseen AVIRIS8 bit40.8023.961.011

Compared with the 32-bit setting, the 8-bit model shows PSNR decreases of 0.12 dB on WHU-Hi and 0.41 dB on unseen AVIRIS, while maintaining comparable RMSE and SAM. It also outperforms the competing 32-bit baselines in Table 2 across all three metrics on both datasets. These results support the effectiveness of QuHIC under low-bit quantization in these additional evaluations.

Manuscript revisions: We have updated Table 2 and p. 8 (Lines 422-423) with the 8-bit results and added corresponding visual comparisons in Figure 11, with the accompanying analysis in Appendix D.8 on p. 21 (Lines 1093–1095).

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

This paper proposes QuHIC, an HSI codec framework that combines lightweight learnable quantized encoding with quantization-consistent multi-prior unrolling. The entropy coding stage uses Huffman coding. The experimental results are extensive, but they lack R-D curves, and comparisons with other methods are limited.

Critical strengths

The paper is well written, with extensive experimental results and a clear structure. The method has good interpretability, employing a learnable lightweight encoder and an unrolled decoder constrained by quantization consistency and multi-prior regularization.

Critical weaknesses

The reported “bitrate” is not the actual entropy-coded bitstream rate, despite the method explicitly using Huffman coding. The bpppb values in Table 1 do not reflect the impact of Huffman coding; therefore, they are insufficient for comparing actual compression efficiency. Table 1 shows 7.7% fewer parameters but 55.1% more FLOPs than BTC-Net V2, which does not seem to be a good trade-off between model size and computational cost. Figure 1 reports quantization bit widths rather than actual bitrates, making the compression efficiency comparison unclear.

Questions for authors
  1. What are the actual bpppb values after Huffman coding?
  2. Could you provide R-D curves using actual bpppb for QuHIC and the compared methods?
  3. Figures 4 and 5 do not specify the representation bit widths or bpppb for the compared methods. Which bit widths and bitrates are used for each method in Figures 4 and 5?
5 comments
AuthorsOpenReview ↗

Response to Reviewer zYhq (1/2)

Thank you for your constructive review and for recognizing the clarity of the framework and the breadth of the experiments. We have provided detailed responses to each of your concerns below. Revisions in the manuscript addressing the questions below are highlighted in blue. For reference, the original submission with a nine-page main text is also provided in the supplementary material.

R1: Actual bitrate after Huffman coding

W1 & Q1: What are the actual bpppb values after Huffman coding?

Response: We apologize for the ambiguity in the original bitrate reporting. We have now updated the low-bit entries to report the actual complete-bitstream bitrates.

1.Full-bitstream bitrate: We measure the complete bitstream size from the actual encoded files, including the Huffman-coded payload, side information required for decoding, and byte-alignment padding. For an HSI with dimensions C×H×WC\times H\times W, let BB denote the total size of the complete bitstream in bytes. The resulting bitrate is R=8BCHWR=\frac{8B}{CHW} bpppb. Huffman coding losslessly preserves the quantized symbols and therefore does not alter the quantized representation supplied to the reconstruction network or the corresponding reconstruction metrics. The updated results for the 8-, 7-, and 6-bit settings are listed below:

Tab. 1. Complete bitstream sizes and bitrates.

Quantization bit depthComplete bitstream size (bytes)Full Huffman-coded bitrate (bpppb)
8 bit91,8030.065
7 bit79,5940.056
6 bit69,2420.049

2.32-bit setting and low-bit settings: The 32-bit setting retains the floating-point latent representation to compare reconstruction capabilities with other floating-point baselines without introducing low-bit uniform quantization. We therefore did not further evaluate the Huffman coding gains for floating-point byte streams. For the low-bit settings, we report the actual bitrate of the complete Huffman-coded bitstream.

Manuscript revisions: We have updated the low-bit bitrates in Table 1 and the corresponding descriptions in Sec. 4.1 on p. 7 (Lines 373–377), and added the complete-bitstream calculation and side-information accounting in Appendix D.4 on p. 17 (Lines 916–917) and p. 18 (Lines 930–949). The 32-bit entries are explicitly identified as uncoded floating-point references on p. 7 (Lines 374–376) and in Appendix D.4.

R2: Rate–Distortion curves

Q2: Could you provide R-D curves using actual bpppb for QuHIC and the compared methods?

Response: Thank you for this valuable suggestion. We have added rate–distortion curves based on actual entropy-coded bpppb in Figure 5 of the revised manuscript.

1.Comparison protocol: The curves compare QuHIC with BTC-Net V2, which supports the evaluated low-bit quantization settings, using actual bpppb on the horizontal axis and PSNR, RMSE, and SAM on the vertical axes. For QuHIC, the bitrate is calculated from the complete encoded bitstream, including decoding side information and byte-alignment padding.

2.Results: Across the overlapping bitrate range, the curves indicate that QuHIC generally achieves higher PSNR and lower RMSE and SAM at comparable actual bitrates, supporting its improved rate–distortion performance.

Manuscript revisions: The rate–distortion curves have been added to Figure 5, with the corresponding analysis in Sec. 4.2 on p. 8 (Line 431), and p.9 (Lines 468-470) of the revised manuscript.

R3: Bitrate annotations in Figures

W1 & Q3: Figure 1 reports quantization bit widths rather than actual bitrates, making the compression efficiency comparison unclear. Figures 4 and 5 do not specify the representation bit widths or bpppb for the compared methods. Which bit widths and bitrates are used for each method in Figures 4 and 5?

Response: Thank you for pointing out this ambiguity.

Manuscript revisions: We have replaced the bit-depth annotations in Figure 1 with bpppb values and added bpppb labels below each method in Figure 4, Figure 6 (formerly Figure 5), Figure 12 (formerly Figure 9), and Figure 13 (formerly Figure 10). For these qualitative comparisons, all methods are evaluated under a common 32-bit floating-point representation setting, corresponding to an uncoded representation rate of 0.314 bpppb.

AuthorsOpenReview ↗

Response to Reviewer zYhq (2/2)

R4: Computational trade-off

W1: Table 1 shows 7.7% fewer parameters but 55.1% more FLOPs than BTC-Net V2, which does not seem to be a good trade-off between model size and computational cost.

Response: Thank you for bringing this issue to our attention. To clarify the trade-off, we have evaluated additional five- and seven-stage variants and separately profiled the encoder and decoder.

1.Reconstruction quality vs unrolling depth: We evaluate the five-, seven-, and nine-stage variants under the same 32-bit setting, keeping the encoder architecture and latent representation size unchanged.

Tab. 2. Complexity and reconstruction quality under the 32-bit setting.

ModelStagesParams (M)FLOPs (G)PSNR (dB) ↑RMSE ↓SAM (°) ↓
BTC-Net V2-12.550565.0341.3948.361.574
QuHIC56.712495.0541.8246.011.462
QuHIC79.234692.4042.1145.031.416
QuHIC911.580876.5342.5942.881.353

Compared with 32-bit BTC-Net V2, five-stage QuHIC uses 46.5% fewer parameters and 12.4% fewer FLOPs, while improving PSNR by 0.43 dB and achieving lower RMSE and SAM. The seven-stage variant reaches 42.11 dB with an intermediate computational cost, while the full nine-stage model achieves the highest reconstruction quality. These results provide alternatives for different reconstruction budgets.

2.Separate encoder and decoder costs: The encoder has the same complexity across all stage configurations: 0.204 M parameters and 1.674 G FLOPs. The decoder is provided below:

Tab. 3. Separate encoder and decoder network complexity.

StagesEncoder Params (M)Decoder Params (M)Encoder FLOPs (G)Decoder FLOPs (G)
50.2046.5081.674493.375
70.2049.0301.674690.724
90.20411.3761.674874.853

Increasing the unrolling depth adds computation to the reconstruction decoder while leaving the encoding cost unchanged. This reflects our intended deployment design: lightweight neural encoding at the acquisition side and more intensive reconstruction at the ground side.

Manuscript revisions: We have added the five- and seven-stage 32-bit results to Table 1, with the corresponding discussion on p. 7 (Line 377) and p. 8 (Line 418). Their reconstruction visualizations are included in Figure 10, and the encoder–decoder complexity breakdown and corresponding analysis are provided in Table 7 and Appendix D.9 on p. 21 (Lines 1099–1104).

Reviewer_zYhqOpenReview ↗

Thank you for the response.

Thank you for the response. However, the R-D curve comparisons are still limited. The BTC-Net V2 results at 7, 6, 5, and 4 bits are taken from the original paper rather than reproduced by the authors. It is unclear how the corresponding actual bitrates in the new curves were obtained. The trade-off between quality and computational cost at low bitrates is also not fully evaluated. These results are not convincing enough, so I will keep my scores.

Reviewer_E5TBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

QuHIC compresses hyperspectral images with a lightweight encoder and reconstructs them with a multi-stage decoder. Each stage combines quantization-interval correction (QCM) with a learned spectral–spatial prior (MPRN). Results on AVIRIS, WHU-Hi, and unseen AVIRIS regions show better PSNR, RMSE, and SAM at several quantization precisions.

Critical strengths

Reconstruction quality improves at the same nominal rate. In the AVIRIS 8-bit setting, Table 1 reports 41.97 dB average PSNR, compared with 40.93 dB for BTC-Net V2. RMSE and SAM also improve. Table 5 reports a 1.81 dB gain at 4-bit. The results cover several datasets and bit settings.

The encoder–decoder split fits the application. Reusing the encoder to check intermediate reconstructions gives quantization information a role throughout decoding. A light airborne encoder and a more expensive ground-side decoder are a reasonable design for this use case. Figures 2–3 make the pipeline easy to follow, and the interval projection and squared-distance penalty are mathematically reasonable.

Critical weaknesses

1. The contribution and consistency claims of QCM are not fully established. The main ablation does not isolate QCM. The A0→A1 step in Table 3 adds deep unrolling and QCM together. Its 8.67 dB gain could come from the multi-stage network, interval correction, or both. A matched comparison with no QCM and with correction toward the quantization center would better show what the interval mechanism contributes.

This matters for novelty. Eq. (3) of D³ (CVPR 2016) already uses the same form of squared penalty outside quantization intervals, and QCS-SGM (ICLR 2023) combines quantized observations with learned priors. QuHIC differs in its representation and full architecture, but needs to show the benefit of its particular combination.

The consistency claims are also stronger than the update supports. The exact gradient in Eq. (19) includes a Sigmoid derivative that Eq. (6) omits. The appendix (p.14) identifies this as an STE approximation, so using the approximation is not itself an error. It does mean that a decrease in the original Eq. (3) is not guaranteed.

Eq. (4) is also a soft penalty, and MPRN can move the re-encoded output outside the interval after QCM. Claims of explicitly enforcing consistency need either direct measurements or more careful wording.

2. The rate and cost comparison is incomplete. The reported rate is not the full bitstream rate. Appendix D.3 says Table 1 uses a fixed-length representation rate without entropy coding. The method section (p.4) describes a Huffman-coded bitstream. The table's bpppb should therefore not be read as the actual transmission cost without the full coded size and side information.

FLOPs rise from 565.03 G for BTC-Net V2 to 876.53 G for QuHIC, about 55%. The reconstruction gains are clear, but the rate–quality–compute trade-off is incomplete. Figure 6 also puts a dataset-average PSNR next to a single-scene image: for example, 41.97 dB at 8-bit rather than the scene's 40.80 dB in Table 1. Please label these values consistently.

Overall assessment

My recommendation is Weak Reject. The reconstruction gains are useful, but the main experiment does not separate interval correction from general multi-stage reconstruction. Clarifying that contribution, the consistency claim, and the actual bitrate would address the reasons for my score. More datasets alone would not answer these questions.

Basis for confidence

I checked the method, equations, tables, appendix, and related work on deep unrolling and quantized reconstruction. I am fairly confident about the comparison and mathematical issues. I did not have the original training code, weights, or full outputs to reproduce the results.

Questions for authors
  1. With the encoder, MPRN, stages, and training setup fixed, how do no QCM, center-point residual correction, and interval residual correction compare? The A0→A1 result does not separate them.

  2. How much do re-encoded outputs violate their intervals after each stage and after the final MPRN? Is Eq. (6) intended as an empirical surrogate update rather than a guaranteed descent step for Eq. (3)?

  3. What is the full size of the Huffman-coded bitstream, including side information, and the resulting bpppb? How should the 32-bit floating-point setting be compared with the low-bit uniform-quantization settings?

  4. Please separate encoder and decoder costs, and confirm whether Figure 6 should show per-scene or dataset-average PSNR.

5 comments
AuthorsOpenReview ↗

Response to Reviewer E5TB (1/5)

Thank you for your thoughtful comments, constructive feedback, and for acknowledging the merits of our work. We will address each of them in turn below. Revisions in the manuscript addressing the questions below are highlighted in blue. For reference, the original submission with a nine-page main text is also provided in the supplementary material.

R1: QCM ablation and the contribution of interval-based correction

W1 & Q1: With the encoder, MPRN, stages, and training setup fixed, how do no QCM, center-point residual correction, and interval residual correction compare? The A0→A1 result does not separate them.

Response: We sincerely apologize for this oversight and thank you for this valuable suggestion. We have conducted two sets of experiments: a comparison of three QCM settings and an additional ablation of the basic unrolling baseline. The details are provided below:

1.Comparison of three QCM settings: We conduct the comparison on the AVIRIS test set under 8-bit quantization, keeping the encoder and MPRN architectures, nine-stage unrolling structure, training data, and optimization settings unchanged. Only the QCM operation is varied:

  • No QCM: The QCM residual correction is disabled, while initialization, the MPRN at each stage, and stage fusion are retained.
  • Center-point QCM: The correction target is the quantization center y~\tilde{y}, using the center-point residual fpred−y~f_{\mathrm{pred}}-\tilde{y}.
  • Interval QCM (full QuHIC): The interval residual of the original model is used. The center-point and interval variants employ the same back-projection structure and step-size learning scheme.

Tab. 1. Comparison of three QCM settings under otherwise identical configurations.

IDQCM settingPSNR (dB) ↑RMSE ↓SAM (°) ↓
Q0No QCM39.7654.041.846
Q1Center-point QCM41.4345.981.525
Q2Interval QCM (full QuHIC)41.9743.981.423

Interval QCM improves PSNR by approximately 2.21 dB over no QCM and 0.54 dB over center-point QCM. The former demonstrates the contribution of residual correction within the complete multi-stage framework, while the latter further supports the benefit of using a quantization interval rather than a single center-point target. The rationale is that a quantization symbol specifies a feasible interval, rather than the exact value before quantization.

2.Basic unrolling baseline between A0 and A1: We introduce A0.5: a basic unrolling network without QCM between A0 and A1. This variant adds a basic multi-stage unrolling structure to A0, with all other configurations matched to the original A1.

Tab. 2. Ablation with the additional basic unrolling baseline A0.5.

IDConfigurationPSNR (dB) ↑RMSE ↓SAM (°) ↓
A0Basic compression–backprojection baseline29.77181.835.018
A0.5Basic unrolling network without QCM36.1480.052.607
A1Basic unrolling network + interval QCM37.5567.712.225

Introducing the unrolling structure increases PSNR to 36.14 dB, and adding QCM further improves it to 37.55 dB. This intermediate baseline A0.5 clarifies the sources of the overall A0→A1 improvement.

3.Contribution: D³ uses a squared penalty for JPEG DCT coefficients outside their quantization intervals, while QCS-SGM combines a pretrained score-based prior with quantization-aware posterior sampling. QuHIC uses interval residuals in the learned latent space to guide stage-wise reconstruction with HSI-specific spectral and spatial priors. Our contribution lies in this reconstruction design within a learned compression framework, rather than a new interval penalty. The Q0–Q2 comparisons above isolate the contribution of interval-based correction within this framework.

Manuscript revisions: We have updated Table 3 to include A0.5 and the three QCM comparisons Q0–Q2. We have also moved the ablation visualization to the main text and added ablations of the three QCM settings, as shown in Figure 7. The corresponding analysis has been revised in Sec. 4.3 on p. 9 (Lines 477–481) and p. 10 (Lines 525–531), and in Appendix D.7 on p. 19 (Line 1024), p. 20 (Line 1079), and p. 21 (Lines 1088–1090).

AuthorsOpenReview ↗

Response to Reviewer E5TB (2/5)

R2: Quantization-interval violations and stage-wise consistency

W1 & Q2: How much do re-encoded outputs violate their intervals after each stage and after the final MPRN?

Response: Thank you for this valuable suggestion. We have added stage-wise results and revised our statements on quantization consistency accordingly.

1.Stage-wise results: We evaluated all nine internal stages of the same complete 9-stage, 8-bit model on the AVIRIS test set. At each stage, we re-encode four states: before QCM, after QCM, the MPRN output before the fusion in Eq. 7 of the paper, and the post-fusion MPRN output. As shown in Tab. 3, we use the intervals with half-width Δ/2\Delta/2 defined in Eq. 2 of the paper. The violation rate is the percentage of re-encoded elements outside their corresponding intervals. The mean squared interval distance is the average, over all latent elements, of each element’s squared shortest distance to its interval, with in-interval elements contributing zero.

Tab. 3. Stage-wise quantization consistency results. Each cell reports violation rate (%) / mean squared interval distance.

StageBefore QCMAfter QCMMPRN before fusionMPRN after fusion
197.72 / 1.06×10−21.06\times10^{-2}96.95 / 3.55×10−33.55\times10^{-3}99.45 / 7.65×10−27.65\times10^{-2}94.99 / 2.72×10−32.72\times10^{-3}
294.99 / 2.72×10−32.72\times10^{-3}91.22 / 4.21×10−44.21\times10^{-4}99.61 / 9.96×10−29.96\times10^{-2}89.61 / 5.51×10−45.51\times10^{-4}
389.61 / 5.51×10−45.51\times10^{-4}50.06 / 1.40×10−51.40\times10^{-5}99.44 / 9.14×10−29.14\times10^{-2}73.46 / 4.77×10−54.77\times10^{-5}
473.46 / 4.77×10−54.77\times10^{-5}23.05 / 7.86×10−77.86\times10^{-7}97.99 / 5.55×10−35.55\times10^{-3}78.99 / 2.90×10−52.90\times10^{-5}
578.99 / 2.90×10−52.90\times10^{-5}32.53 / 6.10×10−76.10\times10^{-7}98.22 / 2.01×10−32.01\times10^{-3}91.00 / 3.01×10−53.01\times10^{-5}
691.00 / 3.01×10−53.01\times10^{-5}48.39 / 6.66×10−76.66\times10^{-7}92.07 / 3.43×10−53.43\times10^{-5}88.92 / 5.66×10−65.66\times10^{-6}
788.92 / 5.66×10−65.66\times10^{-6}49.13 / 1.48×10−71.48\times10^{-7}93.70 / 1.86×10−51.86\times10^{-5}92.31 / 4.49×10−64.49\times10^{-6}
892.31 / 4.49×10−64.49\times10^{-6}40.29 / 9.42×10−89.42\times10^{-8}62.57 / 3.64×10−63.64\times10^{-6}0.33 / 6.06×10−96.06\times10^{-9}
90.33 / 6.06×10−96.06\times10^{-9}0.03 / 3.27×10−113.27\times10^{-11}16.66 / 2.94×10−72.94\times10^{-7}0.29 / 3.56×10−93.56\times10^{-9}

2.Conclusions: QCM reduces both metrics at every stage, supporting its role in improving quantization consistency. The changes from the QCM output to the pre-fusion MPRN output show that refinement can reintroduce interval violations. We therefore designed the fusion operation in Eq. 7 of the paper, which combines the QCM-corrected estimate and the MPRN candidate to mitigate violations introduced by refinement. Both metrics are lower after fusion than at the pre-fusion MPRN output, indicating that this fusion helps alleviate these violations.

Manuscript revisions: We have revised the description of quantization consistency on p. 2 (Lines 72–74) and p. 4 (Lines 173–174, 188–190) to clarify that QCM promotes quantization consistency through interval-residual correction. We describe the residual correction in Eq. 6 on p. 5 (Lines 237–239). The stage-wise results and conclusions have also been added to Table 5 and Appendix D.3 on p. 17 (Lines 905–912).

AuthorsOpenReview ↗

Response to Reviewer E5TB (3/5)

R3: Surrogate update and descent guarantees

W1 & Q2: Is Eq. (6) intended as an empirical surrogate update rather than a guaranteed descent step for Eq. (3)?

Response: We sincerely apologize for not making this distinction clear in the manuscript. Eq. 6 should be understood as a consistency-motivated surrogate update, rather than a guaranteed descent step for Eq. 3. We clarify the details below:

1.Rationale for the approximation: Using the interval residual eke_k defined in Eq. 17 of the paper, the exact gradient in Eq. 19 is $\nabla_{x_k}\mathcal D = H_\theta^\top \left[\sigma'(H_\theta(x_k)+b_\theta)\odot e_k\right]$. The factor σ′(z)=σ(z)(1−σ(z))\sigma'(z)=\sigma(z)(1-\sigma(z)) becomes very small near saturation, attenuating the corresponding interval-residual components before back-projection. Eq. 6 replaces this factor with an identity surrogate so that the interval residual directly provides the correction signal, while a learnable stage-dependent step size controls the correction strength. This is a deliberate approximation, not an equivalent expression for the exact gradient. It defines the correction used during forward reconstruction and serves a different role from the STE used to backpropagate through the quantizer during training.

2.Validation: Since the Sigmoid derivative varies across latent elements, a learned step size does not generally recover the omitted element-wise derivative factor. Thus, the optimization formulation motivates the update but does not establish a descent guarantee. Tab. 3 in R2 shows that the mean squared interval distance decreases after QCM at all nine stages. These results provide direct empirical support for the effectiveness of the surrogate update in improving quantization consistency.

Manuscript revisions: We describe the identity surrogate used in Eq. 6 on p. 5 (Lines 237–239) and clarify that Eq. 6 is a consistency-motivated surrogate update on p. 15 (Lines 805–806). We have also expanded the explanation in Appendix C.1 on p. 15 (Lines 798–805) to clarify the rationale for the approximation and its distinction from the quantizer STE.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Follow-up response: We sincerely hope you can reconsider your score

Response: Thank you for recognizing the additional QCM controls and the clarified file-size breakdown. We apologize for the remaining ambiguity and clarify both points below. We sincerely hope you can reconsider your score.

1.Checkpoint selection in Table 3: The highest average PSNR on the five test images was used to select checkpoints for all Table 3 variants, including A0.5 and Q0–Q2. Thus, the same selection criterion was applied throughout the ablations.

2.Bitrate accounting for BTC-Net V2: QuHIC and BTC-Net V2 follow the same bitrate-accounting convention, including the Huffman payload, codebook information, decoding metadata, and byte-alignment padding. BTC-Net V2 applies this convention across its 4–8-bit settings, with the payload and codebook sizes varying with bit depth. We verify the released implementation using the authors’ official 8-bit checkpoint and cite the corresponding 4–7-bit bitrate–distortion pairs from Table 1 of their publication. Using officially released checkpoints for reproduction, supplemented by clearly attributed published results under consistent evaluation protocols, is standard benchmarking practice.

We would be very grateful if you could reconsider your rating in light of the additional QCM controls, the verified bitstream comparison, and these clarifications. Thank you again for your careful assessment.

Thank you again for your time and consideration. We would appreciate it if you could take another look at our rebuttal when convenient. We hope our clarifications have addressed your concerns, and we would be happy to provide any further information if needed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.