{"id":"c5f21882-e208-4522-a79c-62f5ee8b5745","arxiv_id":"1908.03245","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"GridDehazeNet, a CNN with trainable preprocessing, attention-based multi-scale grid backbone, and post-processing, reports state-of-the-art PSNR/SSIM on RESIDE SOTS, exceeding GFN by over 7 dB indoor.","lead":"A new convolutional network called GridDehazeNet removes haze from single images by learning its own preprocessing filters, combining multiple scales with attention, and skipping the standard physics-based haze model. It reports large gains over prior methods on the RESIDE benchmark, though the results come from a single run without error bars.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Model selection on SOTS makes the reported wide-margin numbers optimistically biased; no held-out validation or repeated runs is reported for the headline comparison.","rationale":"The paper's central claim is the SOTS comparison, so the condition that must hold is that the reported SOTS PSNR/SSIM are unbiased evaluations of a fixed architecture. Section 4.7 breaks this condition: Table 4 explicitly selects r=3,c=6 because it gives the best SOTS numbers, and no validation partition is described. The same table's monotonic improvement suggests the search was truncated at the largest configuration, so the final number is an upper envelope of the sweep. The paper does provide useful independent evidence: ablations in Table 5, a held-out Sun RGB-D set in Table 1, and qualitative real-image results. Those make it plausible that the margin is real, and they are why I would not reject or mark the paper unverdictable. But selection bias plus lack of error bars and code means the exact 'wide margin' number is not trustworthy as reported. A simple hold-out protocol would settle the question; if the margin persists, the claim is credible. This matches the reader's verdict, so I recommend no change.","tokens_in":11447,"tokens_out":6966,"duration_ms":78745,"concrete_test":"Hold out a validation split before any architecture choice: randomly partition SOTS indoor/outdoor into 250/250, run the Table 4 sweep only on one half, freeze r=3,c=6 (or the best configuration), then evaluate the chosen model and retrained GFN on the untouched half with at least 3 seeds. If the held-out indoor PSNR remains above about 30 dB and the GFN gap remains above 5 dB, the selection-bias concern is resolved; if the gap shrinks to within seed noise, the wide-margin claim should be downgraded to conditional.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sections 4.3 and 4.7: the central 'wide margin' claim is quantified on SOTS (indoor 32.16 PSNR vs GFN 24.91), but the same SOTS set is used to choose the backbone geometry. Table 4 sweeps r=1..3 and c=2..6 and the chosen r=3,c=6 is simply the largest configuration, described as yielding the best SOTS numbers; no validation split, no held-out set, and no seed/error-bar information is reported. The reported score is therefore the maximum over the configurations tried, not an unbiased estimate for a pre-specified model, so the margin over GFN is at least partially a selection artifact. The Sun RGB-D rows provide some independent evidence that the method generalizes, but the headline SOTS comparison itself is not a clean evaluation. Because the performance difference is very large (about 7 dB), selection bias may not fully explain it, but the paper currently provides no way to separate bias from true effect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GridDehazeNet, an end-to-end trainable CNN for single-image dehazing that avoids the atmosphere scattering model. The architecture has three modules: a trainable pre-processing module that produces 16 learned input feature maps, an attention-based multi-scale backbone built on a grid network with rows and columns, and a post-processing module. Training uses a combination of smooth L1 loss and VGG16 perceptual loss. The authors report large improvements over DCP, DehazeNet, MSCNN, AOD-Net, and GFN on the SOTS benchmark, additional results on synthetic Sun RGB-D data, qualitative results on real-world images, ablations over grid size and network components, and a runtime comparison. They also propose an explanation for why bypassing the atmosphere scattering model can be beneficial, based on the loss surface of the constrained model.","tokens_in":11670,"tokens_out":3770,"duration_ms":44844,"significance":"If the reported SOTS numbers are unbiased, the paper makes a strong empirical claim: a direct-mapping dehazing network outperforms the previous best method (GFN) by roughly 7 dB indoor PSNR and a large SSIM margin, suggesting that physical-model-based inversion is not necessary even for synthetic data. The architecture is a clean combination of known components, and the paper includes extensive ablations and makes code and a project website available. The main performance evaluation is not circular because it uses external test data and external baselines. However, the strength of the claim is substantially weakened by the model-selection procedure and the absence of variance estimates, so the significance is conditional on fixing the evaluation protocol.","major_comments":[{"comment":"The headline result is not a clean evaluation because the backbone geometry is selected on the test set. The paper chooses r=3, c=6 because Table 4 shows that this configuration gives the best SOTS numbers, with no separate validation set or pre-registered architecture choice. Since SOTS is also the test set used for the main comparison in Table 1, the reported 32.16 dB indoor PSNR is the maximum over the configurations tried rather than an unbiased estimate for a fixed model, and part of the reported margin over GFN is a selection artifact. The authors should either fix the architecture before seeing SOTS, use a held-out validation split for hyperparameter selection, or report the selection procedure explicitly and quantify its effect, for example with bootstrap or a small validation-based model selection study.","section":"Section 4.7 (Table 4)"},{"comment":"All quantitative results come from single training runs with no variance, standard deviation, or number of runs reported. Because training is stochastic, the paper cannot support the claim of a 'wide margin' without evidence that the differences are not within run-to-run variability. The authors should report mean and standard deviation over at least three independent training runs for the main comparison, or provide statistical significance tests for the PSNR/SSIM differences.","section":"Section 4.3 (Table 1) and Tables 2-5"},{"comment":"The statement that all data-driven baselines are 'trained in the same way as the proposed one' is not by itself a fair-comparison protocol. Different architectures may require different learning-rate schedules, numbers of epochs, or regularization to reach their best performance, and the paper reports no per-method hyperparameter tuning or validation. The comparison in Table 1 could therefore reflect a training-schedule disadvantage for some baselines. The authors should either tune each baseline on a validation split or justify that the common schedule is appropriate for all methods, and they should report the baseline training details.","section":"Section 4.3 (training protocol)"},{"comment":"The claim that the atmosphere scattering model creates an 'undesirable loss surface' is not supported by the experiment in Table 2. The indirect network differs from the direct network in output parameterization, the number of parameters, and the optimization task, so its lower PSNR/SSIM could be due to underfitting, a harder estimation problem, or a different effective capacity rather than a loss-surface pathology. To make this load-bearing claim, the authors need controlled comparisons that isolate the loss-surface effect, such as matching capacity, training budget, and optimization difficulty, or direct loss-landscape diagnostics.","section":"Section 4.5 (Table 2)"}],"minor_comments":[{"comment":"There are typos in the opening paragraph, including 'Image dahazing' and 'the proposed hazing method', which should be corrected.","section":"Section 1"},{"comment":"The attention weights in Eq. (2) appear to be trainable scalars per channel rather than input-dependent functions of the features. If so, the mechanism is closer to learned weighted fusion than to attention, and the paper should clarify whether the weights are conditioned on the input or fixed after training.","section":"Section 3.2 (Eq. 2)"},{"comment":"The list of derived inputs says 'three from the gamma corrected GC image' twice and gives only one input from the grayscale image, which appears to be a typo; the total should be checked against the stated 16 inputs.","section":"Section 4.6 (Table 3)"},{"comment":"The runtime comparison is presented only as a plot without numeric values or hardware details in the text; reporting the exact average times and the GPU/CPU used would improve reproducibility.","section":"Section 4.8 (Fig. 8)"},{"comment":"The perceptual loss weight lambda is set to 0.04 without any sensitivity analysis; a short study or a reference justifying this value would strengthen the paper.","section":"Section 3.3 (Eq. 6)"},{"comment":"The real-world evaluation is entirely qualitative. A no-reference dehazing metric or a small user study would make the real-world claim more substantiated, though the authors do note that ground truth is unavailable.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is plausible but the evaluation protocol needs substantial strengthening before the reported margin can be taken at face value. The model-selection issue in Table 4 is the most serious concern because it directly affects the headline SOTS comparison. I found no evidence of circularity or misconduct; the paper should be encouraged to release the code and training details. The loss-surface explanation in Section 4.5 is currently speculative and should be framed as such or replaced with controlled experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a solid engineering paper that convincingly beats prior dehazing methods on SOTS, but the reported margin is inflated by selecting architecture hyperparameters on the test set. The core idea is worth taking seriously.\n\nWhat is actually new: the trainable pre-processing module, which replaces hand-selected inputs, and the attention-based multi-scale grid backbone. The components (GridNet, RDB, channel attention) are known, but the integration for dehazing is original, and the ablations show each piece contributes. The paper also questions whether the atmosphere scattering model helps in learned dehazing, and while that argument is speculative, it is a useful hypothesis. Credit is due for reporting on both SOTS and Sun RGB-D, and for comparing against strong baselines.\n\nThe soft spots are real but not fatal. Most important: Table 4 sweeps grid size and picks r=3, c=6 because it gives the best SOTS numbers, so the 32.16 dB indoor figure is the maximum over configurations, not an unbiased estimate. No validation split, no error bars, no multiple seeds. The 7+ dB gain over GFN is large enough that selection bias alone probably does not explain it, and the Sun RGB-D rows give some independent support, but the paper currently does not separate bias from effect. The baseline comparison is also slightly questionable: all methods are trained with the same schedule, which may disadvantage some. The Sec 4.5 experiment, repurposing the network to estimate transmission and atmospheric light, is a clever but indirect test of the loss-surface explanation, and that explanation remains post hoc.\n\nWho this is for: researchers in image dehazing and restoration, and anyone interested in whether physical models constrain learned methods. The paper is clearly written and the claims are presented honestly enough, but the experimental reporting is not clean enough to take the wide-margin claim at face value.\n\nMy recommendation: send it to peer review. Serious referees should ask for a validation split, repeated runs with seeds, and a fairer training schedule for baselines. The architecture and the central comparison are worth the referee time, and the paper can be made solid with those changes.","headline":"Strong architectural result with a real SOTA claim, but the headline numbers are tuned on the test set; still deserves serious refereeing with revisions.","tokens_in":12206,"tokens_out":1650,"would_cite":true,"duration_ms":19635,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A trainable grid CNN dehazes images without the atmosphere scattering model and tops prior methods by a wide margin on the SOTS benchmark.","keywords":["image dehazing","single image dehazing","convolutional neural network","channel-wise attention","multi-scale estimation","grid network","atmosphere scattering model","RESIDE"],"falsifier":"Run the model-selection procedure on a separate validation split, for example a random subset of the RESIDE indoor training set, and then evaluate the chosen GridDehazeNet on SOTS; if its PSNR falls to the level of GFN (around 25 dB indoor), the headline claim of a wide margin would be refuted. Alternatively, evaluate the same frozen model on a large paired real-world dehazing benchmark; if it does not beat GFN there, the claim of real-world superiority would fail.","tokens_in":1794,"feed_emoji":"🌫️","tokens_out":1683,"duration_ms":73377,"temperature":0.7,"pith_summary":"This paper proposes GridDehazeNet, a fully trainable convolutional network that removes haze from a single image. Its central claim is that an end-to-end network that learns its own pre-processing, fuses features across scales on a dense grid with channel-wise attention, and never invokes the atmosphere scattering model can outperform existing dehazing methods by a large margin on the standard SOTS benchmark and produce visually cleaner results on real hazy photos. If true, it would mean that direct learned mapping is a stronger paradigm for synthetic dehazing than the traditional estimate-then-invert approach, and that the physical model can be dropped without sacrificing synthetic performance. The paper also argues that constraining the solution to the structure of the atmosphere scattering model can trap optimization in a poor local minimum even when that model generated the data.","feed_headline":"Dehazing grid CNN skips physics model and tops SOTS benchmarks","feed_subtitle":"A fully trainable network with learned pre-processing and attention-based grid fusion tops both synthetic and real dehazing comparisons.","key_machinery":"GridDehazeNet is a CNN with three modules. A pre-processing module, consisting of a convolution plus a residual dense block, turns the hazy RGB image into 16 learned feature maps, replacing hand-picked variants such as white balance or contrast enhancement. The backbone is a 3 by 6 grid network: each of three rows operates at a different scale, each of six columns is a bridge that up-samples or down-samples between scales with learned convolutions, and the crossings are densely connected, avoiding the bottleneck of encoder-decoder designs. At each fusion point, channel-wise attention weights $a_r^i$ and $a_c^i$ combine the row and column streams as $\\tilde{F}^i = a_r^i F_r^i + a_c^i F_c^i$, letting the network decide how much each scale contributes. A post-processing module symmetric to the pre-processing one removes artifacts, and the network is trained end to end with a smooth L1 loss plus a VGG16 perceptual loss.","core_discovery":"The paper's central empirical discovery is that GridDehazeNet sets a new best result on the RESIDE SOTS benchmark. It reports indoor PSNR of 32.16 dB and SSIM of 0.9836, and outdoor PSNR of 30.86 dB and SSIM of 0.9819, compared with the previous best GFN at 24.91 dB / 0.9186 indoor and 28.29 dB / 0.9621 outdoor. On a new synthetic set built from Sun RGB-D images, it also leads, with 28.67 dB versus GFN's 25.35 dB. The authors argue that a network which bypasses the atmosphere scattering model can beat model-dependent methods even on synthetic images generated by that model, and they attribute the gain to three components: a trainable pre-processing module that produces 16 learned input feature maps, a 3 by 6 grid backbone that exchanges information across scales without a bottleneck, and channel-wise attention that fuses row and column streams.","pith_inferences":["An unstated consequence is that the headline margin may be optimistic, because the grid size $r=3, c=6$ was chosen from SOTS itself (Table 4); re-measuring on a separate validation split would reveal the true generalization gap.","The loss-surface explanation is testable: train the same network with and without the transmission-map output head over several random seeds; if the gap persists across seeds and datasets, the explanation would be supported.","Because the channel-wise attention weights are few and trainable, one could inspect them to see which scales dominate for thick versus thin haze; the paper does not report such an analysis, but the mechanism invites it.","The synthetic benchmark results do not settle the role of the physics model on real images, since the real-world comparison is qualitative only; a paired real haze benchmark would be needed to confirm the advantage there."],"forward_implications":["If the reported SOTS numbers hold, direct end-to-end dehazing without the atmosphere scattering model is a viable, and on this benchmark superior, alternative to physical-model-based inversion.","Trainable pre-processing can replace hand-selected enhancement variants; the paper's Table 3 shows learned inputs beat both raw inputs and derived inputs such as white balance, contrast enhancement, and gamma correction on SOTS.","The grid backbone with channel-wise attention outperforms the pruned encoder-decoder variant and the original GridNet in the paper's own ablations, so each design choice contributes to the final result.","The paper's loss-surface argument implies that incorporating a heterogeneous physical-model component into a CNN can degrade optimization, so model-based constraints should be introduced with care in learned restoration systems."],"supporting_citations":[{"why":"Provides the RESIDE dataset, including the ITS and OTS training sets and the SOTS test set used for all quantitative comparisons.","marker":"[14]"},{"why":"Gated Fusion Network, the previous state-of-the-art baseline that GridDehazeNet is compared against and the source of hand-selected pre-processing variants.","marker":"[27]"},{"why":"GridNet, the grid network architecture adopted and enhanced for the backbone's multi-scale design.","marker":"[7]"},{"why":"Residual dense block used in the pre-processing and backbone modules, providing the growth-rate-based feature fusion.","marker":"[41]"},{"why":"Inspiration for the channel-wise attention mechanism used in feature fusion.","marker":"[40]"},{"why":"Perceptual loss using VGG16 features, one of the two loss terms used to train the network.","marker":"[10]"},{"why":"Real-world hazy image dataset used for qualitative comparison of dehazing results.","marker":"[6]"},{"why":"Sun RGB-D dataset used to generate a new synthetic test set for robustness comparison.","marker":"[35]"}],"fun_headline_variants":["Dehazing CNN skips scattering model, hits new SOTA","GridDehazeNet: attention grid beats physics-based dehazing","Attention-based grid network tops SOTS dehazing benchmarks","No physics model needed: grid CNN sets dehazing record","Multi-scale grid CNN with attention dethrones physics models"],"cache_read_input_tokens":14336,"weakest_assumption_plain":"The central claim depends on the SOTS numbers being an unbiased measure of dehazing performance, but the grid-size hyperparameters were selected by looking at SOTS itself, so the reported wide margin may be inflated by selection bias.","fun_headline_variants_meta":{"raw":{"variants":["Dehazing CNN skips scattering model, hits new SOTA","GridDehazeNet: attention grid beats physics-based dehazing","Attention-based grid network tops SOTS dehazing benchmarks","No physics model needed: grid CNN sets dehazing record","Multi-scale grid CNN with attention dethrones physics models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1260,"prompt_tokens":944,"completion_tokens":316,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":232}},"tokens_in":560,"tokens_out":316,"duration_ms":3861,"temperature":1.0,"reasoning_tokens":232,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:19:08.419725+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the model-selection procedure on a separate validation split, for example a random subset of the RESIDE indoor training set, and then evaluate the chosen GridDehazeNet on SOTS; if its PSNR falls to the level of GFN (around 25 dB indoor), the headline claim of a wide margin would be refuted. Alternatively, evaluate the same frozen model on a large paired real-world dehazing benchmark; if it does not beat GFN there, the claim of real-world superiority would fail.","supporting_citations":[{"cited_title":"Benchmarking single- image dehazing and beyond","cited_arxiv_id":null,"evidence_quote":"Provides the RESIDE dataset, including the ITS and OTS training sets and the SOTS test set used for all quantitative comparisons."},{"cited_title":"Gated fusion network for single image dehazing","cited_arxiv_id":null,"evidence_quote":"Gated Fusion Network, the previous state-of-the-art baseline that GridDehazeNet is compared against and the source of hand-selected pre-processing variants."},{"cited_title":"Residual Conv-Deconv Grid Network for Semantic Segmentation","cited_arxiv_id":"1707.07958","evidence_quote":"GridNet, the grid network architecture adopted and enhanced for the backbone's multi-scale design."},{"cited_title":"Residual dense network for image super-resolution","cited_arxiv_id":null,"evidence_quote":"Residual dense block used in the pre-processing and backbone modules, providing the growth-rate-based feature fusion."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Inspiration for the channel-wise attention mechanism used in feature fusion."},{"cited_title":"Percep- tual losses for real-time style transfer and super-resolution","cited_arxiv_id":null,"evidence_quote":"Perceptual loss using VGG16 features, one of the two loss terms used to train the network."},{"cited_title":"Dehazing using color-lines","cited_arxiv_id":null,"evidence_quote":"Real-world hazy image dataset used for qualitative comparison of dehazing results."},{"cited_title":"Sun rgb-d: A rgb-d scene understanding benchmark suite","cited_arxiv_id":null,"evidence_quote":"Sun RGB-D dataset used to generate a new synthetic test set for robustness comparison."}],"review_version":1}