{"id":"c65dadcc-51a8-46fe-bf0e-8d9b2d7fabc1","arxiv_id":"2607.15045","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A scene-driven neural network orders Hadamard sensing patterns for single-pixel spectral imaging, improving VIS/NIR reconstructions over fixed orderings at low sampling ratios.","lead":"A deep-learning model chooses which Hadamard patterns to measure next in a single-pixel spectral camera, based on early coarse measurements. The approach improves reconstructed image quality at low sampling budgets and works on a real near-infrared testbed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The magnitude-based oracle support (Eq. 10) is never validated against reconstruction quality for the nonlinear solvers; with F1≈0.68, the learned policy may be a dataset-specific heuristic, undermining the transferability and optimality claims.","rationale":"The reader's weakest assumption identifies the oracle support target as the key unvalidated premise. I agree that this is the most load-bearing concern because the entire training signal is built on it. The central empirical claim—that the proposed method outperforms fixed orderings—is supported by the reported numbers, so the concern does not warrant a REJECT; it does, however, justify a CONDITIONAL verdict because the proxy's alignment with the actual reconstruction objective is unknown. If the reconstruction-aware training test shows no improvement, the concern is resolved; if it shows improvement, the current method is suboptimal and the paper's broader claims about optimality and transferability weaken. The lack of error bars, the emulated real testbed, and missing adaptive baselines are additional weaknesses, but they are secondary to the training-signal foundation. The reader already factored these into a CONDITIONAL verdict, so I recommend no change.","tokens_in":18324,"tokens_out":14576,"duration_ms":172526,"concrete_test":"Train the selector on EuroSAT using a reconstruction-aware loss that backpropagates through a differentiable unrolled version of the CE/TV solver to directly maximize PSNR (or SSIM) at δ=20%, keeping the architecture and predetermined block fixed. Compare test PSNR/SSIM against the BCE-trained selector reported in Table 2. If the reconstruction-aware model does not improve PSNR by at least 0.5 dB, the magnitude-based oracle is a valid proxy; if it does, the current training signal is suboptimal and the central claim is only conditionally supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The entire training signal is a binary cross-entropy loss against m_s = ktop(M_r^p |Y|_Σ, k_a), where |Y|_Σ is the per-band aggregated Hadamard magnitude (Sec. IV.B, Eq. 10). This assumes the largest-magnitude coefficients are the most informative for the PnP, TV, and CE solvers used at test time (Sec. V.C). No experiment validates this proxy against the actual reconstruction objective: the ablations in Fig. 9 compare BCE against an L2-fidelity loss (which uses a transmittance penalty, not the full solver) and an ordering MSE, but none measure PSNR of the final CE reconstruction. The reported F1≈0.68 (Sec. V.B, Suppl. XI.B) means the predetermined Zig-Zag block only partially predicts the oracle support; the network may be learning the average spectral energy distribution of EuroSAT rather than a genuinely scene-driven policy. If the oracle is misaligned or unlearnable, the selector is suboptimal, and the claimed gains over fixed orderings could be smaller or absent on out-of-distribution scenes. The real NIR testbed uses only 8 scenes, too few to rule this out. This is load-bearing because the central claim includes 'transferable across domains' and 'consistently outperform[s] conventional fixed Hadamard orderings.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a scene-driven adaptive ordering of Hadamard patterns for single-pixel spectral imaging. Acquisition is split into a fixed predetermined block (Zig-Zag) and a scene-driven block whose support is predicted by a ResNet-based selector from the initial low-frequency measurements. The selector is trained with binary cross-entropy against an oracle support formed by the largest per-band aggregated Hadamard magnitudes. Reconstruction is performed by transpose, plug-and-play, or consensus-equilibrium solvers. Experiments on EuroSAT and ARAD report consistent PSNR/SSIM/SAM gains over fixed orderings, and an eight-scene NIR testbed with a EuroSAT-trained selector is used to claim cross-domain transferability.","tokens_in":18629,"tokens_out":3855,"duration_ms":49886,"significance":"If the claims hold, the work provides a practical way to make Hadamard single-pixel acquisition content-adaptive while preserving DMD-compatible binary patterns and reconstruction-agnostic sensing. The idea is well motivated and the two-stage formulation is clear. The paper ships code, compares against several fixed orderings, and includes supplementary ablations on loss functions and sampling-ratio splits. The main value is in demonstrating a learned, scene-specific selection of Hadamard coefficients that improves over fixed orderings at low sampling budgets and is transferable to a real NIR setup. However, the central evidence for optimality and for real adaptive acquisition is not yet conclusive: the training target is a magnitude-based proxy that is not validated against reconstruction quality, and the real testbed is an offline emulation rather than an online two-stage acquisition.","major_comments":[{"comment":"The training signal is the oracle support m_s = ktop(M_r^p |Y|_Sigma, k_a), i.e., the largest aggregated Hadamard magnitudes. No experiment shows that this proxy is aligned with what the nonlinear solvers (PnP, TV, CE) actually need. The reported F1≈0.68 means the selector only partially recovers this proxy. The ablations in Suppl. IX.A compare loss functions but never evaluate final reconstruction PSNR/SSIM/SAM against the oracle support itself or against a support optimized for the final solver. Without such a comparison, the claimed optimality of the learned policy is not established, and the transferability claim rests on an untested assumption. I ask for an experiment where the oracle support is replaced by (i) a random support of the same size and (ii) a support selected by greedy reconstruction-error minimization for the CE solver, then compare final metrics.","section":"Sec. IV.B, Eqs. (10)-(11); Sec. V.B"},{"comment":"The real testbed experiments are described as 'real-time reconstruction' and 'online operation', but the protocol acquires the complete Hadamard set once and then numerically permutes coefficients according to each ordering. This is an emulation of adaptive sensing on real measurements, not a true two-stage scene-driven acquisition. The claim in the Abstract and Conclusions of validation on 'real test-bed acquisitions' is therefore overstated. Please clarify this explicitly in the main text, and, if feasible, implement the actual two-stage DMD acquisition for at least one sampling ratio to demonstrate the hardware compatibility claim.","section":"Suppl. XI.A; Sec. VI"},{"comment":"The relative predetermined ratio δ_p = k_p/k used for the main results is not reported. Supplementary Fig. 10 shows that reconstruction quality depends strongly on δ_p, with an optimum around 50-80%. Without specifying δ_p for each table entry, the experiments cannot be reproduced or compared across methods. Also, no error bars or confidence intervals are provided; given that the reported gains over the strongest baselines are about 1-2 dB at some sampling ratios, statistical significance over test splits should be established.","section":"Sec. V.A; Tables 2-4; Suppl. X (Fig. 10)"},{"comment":"The comparison is limited to fixed Hadamard orderings and one static learned-pattern baseline. Existing adaptive/learned Hadamard selection methods (e.g., [12], [13], [37]) are discussed in Related Work but not compared experimentally. Since the paper's contribution is specifically scene-driven ordering, a quantitative comparison against at least the nearest adaptive baseline (e.g., magnitude-sorted selection from [37], or the deep-superpixel method [13]) would strengthen the claim that the proposed learned policy is superior to prior adaptive strategies.","section":"Sec. V.D, Related Work II.C"}],"minor_comments":[{"comment":"The metric strings in the caption/text are difficult to parse; for example, 'Cake Cutting 42.82dB, 0.969, 0.0225' appears to exceed the reference PSNR in the first group. Please align the per-scene layout with the listed metrics so readers can verify the comparisons.","section":"Fig. 8 caption/text"},{"comment":"The phrase 'dropout probability parameter set at 90%' is ambiguous: does it mean keep probability 0.9 or drop probability 0.9? Please clarify.","section":"Sec. V.A"},{"comment":"Although the training is called 'end-to-end', the loss in Eq. (11) is a binary cross-entropy against the oracle support, not a reconstruction loss. The term 'end-to-end' is used in multiple places; consider reserving it for the full sensing-reconstruction pipeline or explicitly defining the end-to-end scope.","section":"Sec. IV.B, Eq. (8)"},{"comment":"The F1-score and accuracy of the mask prediction are reported as single numbers with no variability. Since these are part of the evidence for the learned policy, include means and standard deviations over test folds.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the manuscript is well organized. The main revision points are: validating the oracle-support proxy against the actual reconstruction objective, clarifying and/or upgrading the real testbed from offline emulation to true adaptive acquisition, and reporting the predetermined sampling ratio and statistical variability. These are addressable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does one useful thing well: it takes the idea of using the first low-frequency Hadamard measurements to predict which later patterns to measure, and carries it through to a workable, DMD-compatible system that is reconstruction-agnostic and tested on two multispectral datasets plus a real NIR testbed. The reported gains over fixed orderings (Zig-Zag, XY, Cake-Cutting, Sequency) are consistent—roughly 1–2 dB PSNR across sampling ratios and datasets—and the out-of-distribution transfer to the hardware setup, even with only eight scenes, is encouraging. The method is clearly described and the experiments are reasonably thorough for a subfield paper.\n\nThe main soft spot is the training target. The network is trained with a binary cross-entropy loss against an oracle support defined as the top-k largest Hadamard coefficients by per-band magnitude. That is a plausible proxy for linear reconstruction quality, but the paper never checks it against the actual nonlinear solvers (PnP, TV, CE) used at test time. The reported F1 of about 0.68 against the oracle shows the task is only partially learnable, and one can fairly ask whether the learned policy is capturing scene-specific structure or just the average energy distribution of EuroSAT. However, this is not a load-bearing flaw for the central claim. Even if the oracle is not optimal, the proposed ordering still beats fixed orderings across solvers and datasets, and the paper does not overclaim optimality. The concern is addressable in revision by adding a small ablation—e.g., comparing the oracle support to random support, or to a support selected by actual reconstruction error.\n\nOther weaknesses are minor. The simulation tables have no error bars, which matters a bit when the gains are 1 dB. The baseline set omits other adaptive sensing approaches, including the authors' own earlier superpixel-based work, making it hard to locate the method within the broader adaptive SPI landscape. The real testbed is an emulation—full Hadamard data are acquired and then numerically subsampled—but the paper discloses this in Section VI.A, not only in the supplementary, so the reader's note is too harsh there. Eight scenes is a small demonstration but acceptable for a transfer sanity check. Code is promised without a URL or commit hash; they should provide one.\n\nWho is this for? Researchers working on single-pixel imaging, Hadamard orderings, or adaptive sensing in computational imaging. It is an incremental but solid contribution that deserves a serious referee rather than a desk rejection. For peer review, I would ask for the oracle-validation ablation, error bars, and at least one stronger adaptive baseline before acceptance.","headline":"A solid, incremental adaptive-sensing paper with consistent gains over fixed Hadamard orderings; the untested oracle-support training target is a real but addressable weakness, not a fatal one.","tokens_in":19142,"tokens_out":2812,"would_cite":true,"duration_ms":34854,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","94A08","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Scene-driven selection of Hadamard patterns—based on a small initial block of measurements—improves single-pixel spectral image reconstruction compared with fixed orderings, in simulation and on a real near-infrared testbed.","keywords":["Hadamard single-pixel imaging","spectral imaging","adaptive sensing","scene-driven sampling","deep learning","compressed sensing","near-infrared imaging","binary modulation patterns"],"falsifier":"Take a set of small scenes, fully measure their Hadamard spectra, and for each scene exhaustively search all k_a-subsets of unmeasured coefficients to find the subset that actually minimizes reconstruction error for a given solver. If that optimal subset consistently differs from the top-magnitude oracle support m_s, or if reconstructing from m_s is beaten by an equally-sized alternative subset, the paper's central premise fails and the learned selector is chasing the wrong target.","tokens_in":18171,"feed_emoji":"⚙️","tokens_out":4861,"duration_ms":49140,"temperature":0.7,"pith_summary":"This paper sets out to show that a single-pixel spectral imager does not have to rely on a fixed, scene-independent ordering of Hadamard modulation patterns. Instead, after capturing a small predetermined block of low-frequency Hadamard coefficients, a neural network predicts which of the remaining coefficients carry the most scene energy, and those patterns are measured next. The authors claim this scene-adaptive selection consistently beats conventional orderings (Sequency, Zig-Zag, XY, Cake-Cutting) at sampling ratios from 5% to 30% in terms of PSNR, SSIM, and spectral angle, and that the learned selector transfers from satellite imagery to a real NIR testbed without retraining. If correct, this offers a practical path to faster, higher-quality spectral imaging on existing DMD hardware.","feed_headline":"Scene-driven Hadamard selection lifts spectral image quality","feed_subtitle":"A network picks which Hadamard patterns to measure next for each scene, improving PSNR and spectral fidelity.","key_machinery":"The load-bearing mechanism is the two-stage acquisition protocol with a learned support predictor. Stage one selects k_p Hadamard rows via a predetermined ordering (typically Zig-Zag); stage two computes m_a = P_theta(Y_p), masks out already-sensed rows, picks the k_a largest remaining sigmoid scores, and acquires those rows. The training target is the oracle sparse support m_s = top-k of |Y|_Sigma, and the loss is plain binary cross-entropy. This turns 'which patterns next?' into a binary classification problem whose labels come from coefficient magnitude, and it lets the sampling budget be controlled by construction.","core_discovery":"The central claim is that an end-to-end trained binary classifier can convert a small set of low-frequency Hadamard measurements into a scene-dependent ordering of the remaining Hadamard basis, and that this ordering improves reconstruction quality. The network P_theta receives the predetermined measurements Y_p and outputs a support mask m_a over unmeasured coefficients; it is trained with binary cross-entropy against an oracle support m_s defined by the top-k_a remaining coefficients ranked by per-band aggregated magnitude |Y|_Sigma. Because the sensing patterns stay strict Hadamard ±1 rows, the method remains DMD-compatible, and because the ordering is a property of the measurement set ra","pith_inferences":["The paper's oracle target—largest-magnitude remaining Hadamard coefficients—is a proxy for reconstruction value, not a proof of optimality; a loss that measures downstream reconstruction error directly could yield a different and possibly better support.","The two-stage idea extends naturally to multi-stage or video settings: once the first adaptive block is measured, the predictor could be re-applied to the growing set of coefficients to refine the remaining budget.","The approach should carry over to other orthogonal bases (e.g., Fourier or Walsh) for which binary or sign-binary patterns are implementable, since the machinery only requires a fast transform and a magnitude-ranked oracle.","The reported F1 of ~0.68 suggests a ceiling imposed by the target definition; if the field adopted a reconstruction-aware oracle, the same architecture might show larger apparent gains."],"forward_implications":["Fixed Hadamard orderings leave reconstruction quality on the table; a scene-aware selection of the same physical patterns recovers up to roughly 2 dB more PSNR at low sampling ratios.","The learned ordering is reconstruction-agnostic: the same selected mask improves a simple transpose backprojection and sophisticated iterative solvers, meaning hardware gains do not depend on a particular algorithm.","The selector trained on satellite spectral images transfers to a different near-infrared single-pixel camera without fine-tuning, suggesting the learned criterion is about scene statistics rather than dataset-specific details.","Because the method only reorders existing Hadamard patterns, it can be layered onto high-speed cyclic-mask or learned-pattern systems to combine per-pattern speed with compressive efficiency.","Mask-level metrics (accuracy ~0.88–0.91, F1 ~0.68–0.70) show the support predictor is informative but imperfect, so reconstruction quality may improve further with better supervision or richer architectures."],"fun_headline_variants":["Scene-driven Hadamard ordering boosts spectral image quality","AI picks Hadamard patterns to sharpen spectral images","Spectral imaging gets scene-adaptive Hadamard ordering","End-to-end ordering of Hadamard basis for better SPI","Scene-specific Hadamard sequence improves single-pixel spectra"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole training signal rests on the premise that the largest-magnitude remaining Hadamard coefficients, aggregated over spectral bands, are the most informative coefficients to acquire for the reconstruction solvers used—a premise the paper does not prove and that the modest F1 score (about 0.68) suggests is only partially learned.","fun_headline_variants_meta":{"raw":{"variants":["Scene-driven Hadamard ordering boosts spectral image quality","AI picks Hadamard patterns to sharpen spectral images","Spectral imaging gets scene-adaptive Hadamard ordering","End-to-end ordering of Hadamard basis for better SPI","Scene-specific Hadamard sequence improves single-pixel spectra"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1173,"prompt_tokens":694,"completion_tokens":479,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":409}},"tokens_in":438,"tokens_out":479,"duration_ms":4797,"temperature":1.0,"reasoning_tokens":409,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:18:59.589947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of small scenes, fully measure their Hadamard spectra, and for each scene exhaustively search all k_a-subsets of unmeasured coefficients to find the subset that actually minimizes reconstruction error for a given solver. If that optimal subset consistently differs from the top-magnitude oracle support m_s, or if reconstructing from m_s is beaten by an equally-sized alternative subset, the paper's central premise fails and the learned selector is chasing the wrong target.","supporting_citations":[],"review_version":1}