{"id":"d324fc38-7b66-4555-a6a1-2d04d405391f","arxiv_id":"2411.16336","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A wavelet-domain deep unfolding network with adaptive subband sampling and tree-structured reconstruction reports state-of-the-art PSNR/SSIM on Set5, Set11, Set14, and Urban100 at CS ratios 10-50%.","lead":"WTDUN is a deep unfolding network for image compressed sensing that splits images into wavelet subbands, gives more measurements to important subbands, and reconstructs with a tree-structured prior. It reports consistent PSNR/SSIM gains on standard benchmarks, though its adaptive sampling relies on the original image's statistics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WAS allocation uses ground-truth subband statistics (Eq. 6) to choose per-subband measurement counts; at test time this requires an oracle or uncompensated side channel, so the 1.5-2.5 dB adaptive-sampling gain and the SOTA claim are not established for a standard fixed-operator CS task.","rationale":"The paper is a coherent engineering contribution: the wavelet-domain unfolding is clearly described, ablations cover WAS, WTP, deblock, MC, and CAC, and the reported trends are internally consistent. The empirical numbers may well have been produced exactly as described. The question is whether they answer the same problem as the baselines. The strongest claim is SOTA PSNR at every sampling ratio; the largest component of that claim is the WAS module, whose ablation contributes 1.5-2.5 dB. That module requires statistics of the original image (Eq. (6), Section 3.2.1), which are unavailable in one-shot CS and, if made available via an adaptive protocol, must be paid for as side information. The paper does not disclose such side information or compare with baselines under the same adaptive protocol, so the central comparison conflates sensing protocol with reconstruction quality. This is the same weakest assumption the reader identified. A controlled pilot-based or rate-accounted experiment would settle it. No formal verification or released code is provided, but the absence of those does not by itself change the assessment; the missing protocol condition is sufficient to keep the verdict conditional. I therefore recommend UNCHANGED, i.e., the reader's CONDITIONAL verdict remains appropriate.","tokens_in":19181,"tokens_out":6878,"duration_ms":68097,"concrete_test":"Retrain WTDUN with the WAS allocation computed from a causal pilot estimate (e.g., a uniform 2% measurement pass per block, or the LL subband alone) instead of from ground-truth |theta_s| in Eq. (6), keeping the total measurement budget and architecture fixed, and evaluate on Urban100 and Set11 at 20% and 50%. Also serialize the resulting allocation as metadata and include its bit cost in a rate-distortion comparison. If the pilot-based WAS gain over no-WAS is materially below the 1.5-2.5 dB in Tables 6-7, the reported gain depends on oracle side information and the SOTA claim should be restricted to the privileged setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the encoder can evaluate Eq. (6) for the unknown test image before allocating measurements. In Section 3.2.1, the allocation weight is computed from the mean and standard deviation of the absolute wavelet coefficients |theta_s| of the original image X, and those statistics determine the number of measurements M_s per subband. In one-shot compressed sensing the encoder only has y = A x; it cannot know |theta_s| unless the protocol is two-pass or oracle-assisted. If the protocol is adaptive, the per-image or per-block allocation must be transmitted to the decoder as side information, and Tables 6 and 7 report no such overhead. None of the comparison methods is described as receiving original-image subband statistics as side information; they use fixed or learned operators, so WTDUN is evaluated under a different, informationally privileged protocol. Since the WAS ablation alone contributes 1.5-2.5 dB at 20% and 50% ratios (Section 4.3.1), this protocol mismatch, not the wavelet tree prior or the unfolding architecture, is what currently supports the claim of higher PSNR at all sampling ratios. The paper never states how the allocation is obtained at inference or accounts for its bit cost.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WTDUN, a deep unfolding network for image compressed sensing that operates on Haar wavelet subbands. It combines a wavelet-domain adaptive sampling (WAS) module that allocates measurement counts per subband according to subband energy and sparsity, a tree-structured reconstruction prior encoded through group soft-thresholding, and K unfolding stages with deblocking and cross-domain attention. The authors report PSNR/SSIM gains over several recent deep CS methods on Set5, Set11, Set14, and Urban100, and ablations attribute large gains to WAS and to the deblocking module.","tokens_in":1386,"tokens_out":1704,"duration_ms":68806,"significance":"If the reported results are taken at face value and the sampling protocol is a legitimate compressed-sensing protocol, WTDUN is a strong empirical contribution: the tables cover four benchmarks and five sampling ratios, the ablation study isolates the effect of each module, and the WAS ablation cleanly demonstrates a large effect on reconstruction quality. However, two load-bearing issues, namely the self-referential update in Eq. (5) and the oracle-like adaptive sampling protocol in Section 3.2.1, mean that the paper's central claims of state-of-the-art performance and of optimization grounding are not yet established.","major_comments":[{"comment":"The update x^k = A^T z^{k-1} + x^{k-1} - (A^T A - I)(x - x^{k-1}) defines the new iterate in terms of the unknown ground-truth signal x. This cannot serve as a derivation of a computable reconstruction iteration, and it does not provide optimization grounding for the network, because the implemented module replaces the offending term with the learned denoising block N^k. Please replace this derivation with a valid proximal or denoising update, or state explicitly that the network is only inspired by this identity rather than derived from it.","section":"Section 3.1, Eq. (5)"},{"comment":"The WAS allocation rule uses the mean and standard deviation of |theta_s|, where theta_s = Psi X is the wavelet transform of the original image X. In one-shot compressed sensing, the encoder only has y = A x and cannot compute these statistics before choosing the measurement counts M_s. If the intended protocol is adaptive, the per-image or per-block allocation must be transmitted to the decoder as side information, and neither Tables 6-7 nor the comparisons in Tables 1-4 account for this cost; none of the comparison methods is given original-image subband statistics. Since the WAS ablation alone contributes about 1.5-2.5 dB at 20% and 50% ratios, the claim of higher PSNR than fixed-operator or learned-operator baselines at all sampling ratios is not supported by the current experiments. Please specify the inference-time protocol, report the side-information overhead, or evaluate under a fixed measurement operator.","section":"Section 3.2.1 and Section 4.3.1"},{"comment":"The text says the denoising block N^k(theta^{k-1}) removes noise from theta^{k-1}, but Eq. (18) and the surrounding derivation require this term to play the role of the unknown error (x - x^{k-1}) from Eq. (5). These are different objects, and a CNN block operating on the current estimate does not obviously compute either one. Please clarify the role of N^k; if it is meant to approximate the error term, explain how it is trained to do so.","section":"Section 3.3.2, Eq. (18)"}],"minor_comments":[{"comment":"The index condition 1 < i < L is inconsistent with Eq. (6), where 1 <= i <= l; it should presumably be 1 <= i <= L or the notation should be aligned.","section":"Eq. (9)"},{"comment":"The notation for the bias term, written as (n sum_{i=0} bias_i = 0), is garbled; please define bias_i and its index range explicitly.","section":"Section 3.2.1"},{"comment":"The text says the control parameter alpha is initialized as 1, but Section 3.2.1 introduced the allocation control parameter eta; please clarify whether alpha and eta are the same parameter or different ones.","section":"Section 4.1"},{"comment":"Several entries for DPA-Net are missing (denoted by *), yet Section 4.2.1 claims higher PSNR at all sampling ratios; please state explicitly that incomplete comparisons are excluded from that claim.","section":"Tables 1 and 2"},{"comment":"The soft-threshold formula divides by the norm of G theta^{k-1}; please specify the behavior when this norm is zero, for example by defining the expression to be zero in that case.","section":"Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The empirical tables are extensive and the ablation study is informative, but the fair-protocol issue is central. If the authors can provide experiments with a fixed measurement operator or with side-information cost fully accounted for, the contribution could be solid. The paper does not mention code release, which limits reproducibility but is not by itself a reason for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Actionable summary: this is a solid deep-unfolding architecture for image CS with one serious protocol flaw that needs fixing before the SOTA claims can land.\n\nWhat's good: The authors build a wavelet-domain deep unfolding network (WTDUN) that operates on multi-scale subbands. Each stage has a gradient descent step, a group soft-thresholding derived from a tree-structured group sparsity prior, full-image deblocking, and a cross-domain attention block before the inverse wavelet. The ablation study is genuinely careful: they isolate WAS, WTP, deblock, memory context, and cross-attention, and report numbers on two datasets at two ratios. Experiments cover four datasets, six sampling ratios, and a dozen learned baselines plus model-based multi-scale methods. If the reported numbers are taken at face value, WTDUN has a real edge, especially on Urban100 at higher rates.\n\nWhere it gets soft: the edge appears to come from an informationally privileged sampling rule. Section 3.2.1 computes the allocation weight from the standard deviation and mean of the absolute wavelet coefficients |theta_s| of the original image X. That is ground truth. In one-shot CS, the encoder only has y = Ax; it cannot know |theta_s| without a pilot measurement or side information. The paper does not describe how this allocation is obtained at inference nor does it budget the cost of side information. The ablation credits WAS with 1.5-2.5 dB, so this issue is not peripheral. The comparison methods are not given this advantage. The stress-test note is right, and I don't think the paper has an answer in its current form.\n\nAlso, Eq. (5) as written uses the unknown x on the RHS to define x^k. In practice the implementation replaces the oracle term with a learned denoiser, which is what any deep-unfolding paper should do, but the text needs to say that rather than presenting a circular derivation.\n\nThe rest is fine. The complexity analysis is standard. The writing is readable. No code release in the arXiv version, so reimplementation is needed to verify, but the experimental effort already shown is substantial.\n\nThis paper should go to peer review. It is not ready in this form, but the architecture is plausible and the protocol question is fixable - state the inference-time allocation (e.g., from an initial estimate) or add the side-channel cost, rerun the comparisons under a matched protocol, and fix the Eq. (5) derivation. I'd then be willing to accept it as a solid contribution, even if not a breakthrough. For a reading group, it's a good case study in how benchmarking protocols can quietly give a method an unfair advantage.\n\nWho is this for: people working on deep unfolding for CS, adaptive sampling, and wavelet-based image reconstruction. They should be aware of the current claims and check the allocation question before citing.","headline":"A well-engineered wavelet-domain deep unfolding system for image CS whose headline gains rest on an unstated oracle advantage in the adaptive sampling rule; fix that and the derivation slip, and it is a solid paper.","tokens_in":20056,"tokens_out":8852,"would_cite":true,"duration_ms":87948,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"WTDUN claims that organizing both sampling and reconstruction around the wavelet coefficient tree lifts image compressed sensing to higher PSNR than prior deep-learning methods at every tested sampling ratio.","keywords":["compressed sensing","deep unfolding","wavelet transform","adaptive sampling","tree-structured sparsity","image reconstruction","multi-scale subbands"],"falsifier":"Run WTDUN on Urban100 at 50% sampling with the per-subband measurement allocation replaced by a uniform allocation of the same total budget while keeping every other component and the training procedure fixed; the paper's own ablations predict a PSNR drop of about 1.5–2.5 dB, so a result showing parity or a uniform-allocation win would refute the adaptive-sampling claim. Independently, reproducing the reported absolute PSNR numbers on the four benchmarks with the paper's protocol would confirm or refute the headline performance claim.","tokens_in":18979,"feed_emoji":"🖼️","tokens_out":8259,"duration_ms":73530,"temperature":0.7,"pith_summary":"The paper tries to show that image compressed sensing improves when both sampling and reconstruction are organized by the multi-scale wavelet structure of an image instead of treating all pixels or all frequency bands alike. Its method, WTDUN, does two linked things: it assigns the limited measurement budget across wavelet subbands according to each subband's energy and sparsity, and it unfolds an iterative optimization into a deep network whose steps are guided by the parent-child tree structure of wavelet coefficients. On Set5, Set11, Set14, and Urban100, the 128×128 variant of WTDUN reports higher PSNR than every compared deep-learning method at all tested sampling ratios, with gains of 0.20–2.22 dB at 10% and 1.01–5.57 dB at 50% on Urban100. If correct, this says that frequency-aware adaptive sampling plus cross-scale reconstruction is the right design direction for practical compressed-sensing encoders.","feed_headline":"Wavelet tree sampling lifts compressed-sensing images by up to 5.57 dB","feed_subtitle":"Adaptive per-subband measurement allocation plus tree-structured deep unfolding beats prior deep CS methods at every tested ratio.","key_machinery":"The carrying object is the wavelet coefficient tree. A multi-level Haar transform splits each image block into a low-frequency LL subband plus high-frequency subbands at several scales, and WTDUN treats those subbands as leaves of a quadtree. In sampling, the standard deviation and mean of each subband's absolute coefficients are weighted to set the number of measurements that subband receives (Eq. 6 and the allocation algorithm of Section 3.2.1), so low-frequency energy and high-frequency texture each get a proportionate share of the budget. In reconstruction, the tree-structured prior expresses the structural sparsity of the quadtree: coefficients in parent-child groups are encouraged to shrink together through a group-wise soft threshold (Eq. 12), and an unfolded gradient descent module alternates with denoising and deblocking. The tree is what lets the network transfer information between scales, which the paper credits for its improved texture and edge recovery.","core_discovery":"The central discovery is that wavelet-domain adaptive allocation of measurements is a large, mostly untapped source of reconstruction gain in learned compressed sensing. The paper's ablation attributes about 1.5–2.5 dB of PSNR on Urban100 and Set11 to the subband-importance allocation, about 0.08–0.2 dB to the wavelet tree-structured prior, and about 1–2.5 dB to full-image deblocking. The full architecture—initial pseudo-inverse reconstruction, nine unfolded stages alternating gradient descent with group-wise soft shrinkage on parent-child coefficient groups, full-image deblocking, and cross-domain attention—achieves the best numbers in the comparison tables on all four datasets at every tested sampling ratio, with the largest margins at the 50% ratio where the measurement budget is most flexible.","pith_inferences":["A practical consequence the paper does not develop: the adaptive allocation depends on the original image's own wavelet statistics, so a real single-shot sensor would need to transmit the allocation table as side information or estimate it from a coarse preview; the reported gains should be re-measured once that overhead is charged to the measurement budget.","The same wavelet-tree prior could be transplanted into other deep unfolding networks; since the paper's ablation shows the prior's independent gain is modest, a fair comparison would hold the sampling protocol fixed and swap only the reconstruction prior.","The large margin at 50% over uniform-sampling baselines suggests that most remaining headroom in block compressed sensing lies in how measurements are spent, not just in the reconstruction network, which generalizes the result beyond this specific architecture."],"forward_implications":["If the reported numbers hold, WTDUN achieves the highest PSNR on these block-based compressed-sensing benchmarks at 10–50% sampling ratios, with the largest margins at higher ratios.","The ablation isolates where the gain comes from: adaptive subband allocation accounts for roughly 1.5–2.5 dB, so future learned CS encoders should treat measurement allocation as a first-class design variable rather than a fixed random matrix.","The wavelet tree-structured prior, though smaller in effect at about 0.08–0.2 dB, is a checkable mechanism for recovering fine textures and can be added to other unfolding networks without changing their sampling setup.","Full-image deblocking contributes about 1–2.5 dB, meaning block-based methods that ignore inter-block context are leaving a large part of achievable quality on the table.","Two-level wavelet decomposition with 64×64 blocks beats one-level and three-level decomposition, so the hierarchy depth matters: too shallow loses sparsity, and too deep starves the small subbands of features the network can use."],"supporting_citations":[{"why":"Establishes the ISTA deep-unfolding paradigm that WTDUN builds on and supplies a comparison baseline.","marker":"[51]"},{"why":"Provides the denoising-perspective update equations (2)–(5) and the full-image deblocking idea that WTDUN adapts, and is a key baseline.","marker":"[56]"},{"why":"The classical multi-scale block compressed-sensing algorithm in the wavelet domain that motivates subband-structured sampling and serves as a Table 5 baseline.","marker":"[15]"},{"why":"Memory-augmented deep unfolding baseline that WTDUN compares against and from which the memory-context idea in its deblocking stages derives.","marker":"[34]"},{"why":"Flexible ISTA-Net++ unfolding baseline compared throughout, including in the phase-count study of Table 10.","marker":"[48]"},{"why":"Block compressed sensing with wavelet and directional transforms, a classical baseline for the multi-scale sampling comparison in Table 5.","marker":"[28]"},{"why":"Optimization-inspired compact deep CS network (OPINE-Net) whose learned multi-scale sampling is a direct competitor on Urban100.","marker":"[53]"},{"why":"Supplies the cross-domain attention (CAC) module that WTDUN uses before the inverse wavelet transform to fuse low- and high-frequency subbands.","marker":"[44]"}],"fun_headline_variants":["Wavelet tree sampling adapts measurements per subband for sharper CS images","Deep unfolding on wavelet trees gives adaptive CS sampling gains","Adaptive wavelet subband allocation boosts compressed sensing quality","Tree-structured wavelet unfolding outperforms uniform CS methods","Multi-scale wavelet sampling lifts compressed sensing PSNR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The performance figures presuppose that the encoder may look at the original image's wavelet subband statistics and spend a different number of measurements on each subband of each image, with the decoder knowing exactly how the budget was spent; if compressed sensing must use one fixed measurement operator independent of content, or if the allocation costs side information that must fit inside the same bit budget, the advertised 1.5–2.5 dB from adaptive sampling will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet tree sampling adapts measurements per subband for sharper CS images","Deep unfolding on wavelet trees gives adaptive CS sampling gains","Adaptive wavelet subband allocation boosts compressed sensing quality","Tree-structured wavelet unfolding outperforms uniform CS methods","Multi-scale wavelet sampling lifts compressed sensing PSNR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1579,"prompt_tokens":985,"completion_tokens":594,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":515}},"tokens_in":601,"tokens_out":594,"duration_ms":6230,"temperature":1.0,"reasoning_tokens":515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:14:58.571609+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run WTDUN on Urban100 at 50% sampling with the per-subband measurement allocation replaced by a uniform allocation of the same total budget while keeping every other component and the training procedure fixed; the paper's own ablations predict a PSNR drop of about 1.5–2.5 dB, so a result showing parity or a uniform-allocation win would refute the adaptive-sampling claim. Independently, reproducing the reported absolute PSNR numbers on the four benchmarks with the paper's protocol would confirm or refute the headline performance claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the ISTA deep-unfolding paradigm that WTDUN builds on and supplies a comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the denoising-perspective update equations (2)–(5) and the full-image deblocking idea that WTDUN adapts, and is a key baseline."},{"cited_title":"Fowler, Sungkwang Mun, and Eric W","cited_arxiv_id":null,"evidence_quote":"The classical multi-scale block compressed-sensing algorithm in the wavelet domain that motivates subband-structured sampling and serves as a Table 5 baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Memory-augmented deep unfolding baseline that WTDUN compares against and from which the memory-context idea in its deblocking stages derives."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Flexible ISTA-Net++ unfolding baseline compared throughout, including in the phase-count study of Table 10."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Block compressed sensing with wavelet and directional transforms, a classical baseline for the multi-scale sampling comparison in Table 5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Optimization-inspired compact deep CS network (OPINE-Net) whose learned multi-scale sampling is a direct competitor on Urban100."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the cross-domain attention (CAC) module that WTDUN uses before the inverse wavelet transform to fuse low- and high-frequency subbands."}],"review_version":1}