{"id":"22e1005b-7c49-466d-b568-c786eb8f286c","arxiv_id":"1908.10267","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-subnetwork deraining model, DRD-Net, combines SE-based rain residual learning with a dilated-context detail repair block to improve rain removal and detail recovery.","lead":"DRD-Net is a two-part deep network that removes rain from single images and then repairs image details lost during removal. The paper reports state-of-the-art PSNR and SSIM on three synthetic benchmarks and shows smoother real-world deraining results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2's baseline values contain implausible duplicates (DDN=GMM on Rain800; GMM=DSC in Fig. 1), so the quantitative outperformance claim is not currently reliable.","rationale":"The reader's verdict (CONDITIONAL) is appropriate, but the weakest assumption identified there (synthetic rain model fidelity) is not the most load-bearing issue for the paper's central claim. The strongest claimed evidence is the quantitative comparison in Table 2. That table contains an entry in which DDN and GMM have identical PSNR and SSIM on Rain800, and Fig. 1 lists identical scores for GMM and DSC. These are not subtle differences; they are exact duplicates that two different algorithms should not produce. This suggests the baseline numbers were copied or transcribed incorrectly. Since the claim is that DRD-Net outperforms these methods, incorrect baseline values directly undermine the comparison. The ablation study and released source code are positive signs, but they do not validate the numbers in the comparison table. The correct response is to keep the paper under conditional acceptance: ask the authors to verify and correct Table 2, ideally by re-running all baselines with the same protocol, and then re-evaluate the claimed margins. The synthetic-to-real gap is a legitimate but separate concern that applies to the whole field and does not need to block the benchmark claim. Thus the verdict should remain CONDITIONAL (no change).","tokens_in":14634,"tokens_out":5159,"duration_ms":49815,"concrete_test":"Run the released DRD-Net model and the official implementations of DDN, RESCAN, DAF-Net, GMM, and DSC on the Rain200L, Rain200H, and Rain800 test splits (or use the authors' pre-trained models if provided) and recompute PSNR/SSIM with the same code. Specifically check whether DDN's Rain800 PSNR/SSIM differs from GMM's 24.04/0.8675; if it does, Table 2 contains a transcription error and the claimed outperformance margins need to be recomputed before the comparison can be accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DRD-Net outperforms state-of-the-art deraining methods depends directly on the quantitative results in Table 2 and the qualitative figures. Table 2 reports the DDN entry on Rain800 as PSNR 24.04 / SSIM 0.8675, which is bit-for-bit identical to the GMM entry on the same dataset. Two structurally unrelated algorithms producing identical PSNR and SSIM on the same test set is effectively impossible unless the values were duplicated by mistake. The same problem appears in Fig. 1, where GMM and DSC are both listed as 10.61 / 0.4675 for the same input image. These duplicates indicate that at least some baseline numbers were transcribed or copied incorrectly. Because the claimed margins over the second-best method are small (e.g., 37.15 vs 37.07 on Rain200L), even a small correction to a baseline entry could change the ranking. The paper says all methods were run under the same training and testing datasets but does not state whether the numbers were recomputed with the provided code or taken from the original papers, nor does it provide error bars or repeated runs. Until the baseline numbers are independently verified, the central claim is unsupported. The real-world visual comparisons are suggestive but not quantified, so they cannot rescue the quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DRD-Net, an end-to-end single-image deraining network composed of two sub-networks: a rain residual network that predicts rain streaks using squeeze-and-excitation residual blocks, and a detail repair network built from a newly proposed structure detail context aggregation block (SDCAB) with dilated convolutions, intended to recover image details lost during deraining. The two sub-networks are trained jointly with a weighted L2 loss. The method is evaluated on Rain200L, Rain200H, and Rain800 synthetic benchmarks plus real-world images, with PSNR/SSIM comparisons against GMM, DSC, DDN, RESCAN, DAF-Net, and UGSM, and the paper reports state-of-the-art quantitative results and qualitative detail preservation.","tokens_in":15054,"tokens_out":4339,"duration_ms":43511,"significance":"The proposed architecture is a clear and plausible contribution to the deraining literature: the two-stage design with a dedicated detail-repair network is well motivated, and the ablation study in Table 3 supports the contribution of the SDCAB block. The authors also provide source code on GitHub, which aids reproducibility. However, the central quantitative claim currently rests on a table that contains apparent transcription errors, and the specification of the detail-repair network input is internally inconsistent. If the quantitative results survive verification and the architecture description is corrected, the paper would be a useful contribution to the community.","major_comments":[{"comment":"Table 2 reports identical PSNR/SSIM values for DDN and GMM on Rain800 (24.04/0.8675), and Fig. 1 lists the same PSNR/SSIM for GMM and DSC (10.61/0.4675). Two structurally unrelated baselines producing bit-identical metrics on the same data is effectively impossible unless the values were copied or transcribed incorrectly. Since the reported margin over RESCAN on Rain200L is only 0.08 dB (37.15 vs. 37.07), even a small correction to a baseline entry could change the ranking. The authors must verify every baseline number, either by rerunning with the provided code under identical protocols or by clearly citing the original source of each number, and report whether the values are reproduced or taken from previous papers.","section":"Section 5.2, Table 2 and Fig. 1"},{"comment":"There is an internal inconsistency about the input to the detail repair network. The text says the network operates on the preliminarily derained image Ip, but Eq. (6) trains a function g(·) on the rainy image Oi, and Eq. (10) sets layer0 = PRelu(Conv 3x3(O)), i.e., the network input is the rainy image. If the network takes the rainy image as input, the surrounding description should be corrected; if it takes Ip, then Eqs. (6) and (10) are wrong. This ambiguity materially affects reproducibility of the proposed method and needs to be resolved.","section":"Section 4.2, Eqs. (6) and (10)"},{"comment":"The dilation structure of the SDCAB is described inconsistently. In Section 4.2, SDCAB is said to contain dilated convolutions with dilation scales 1, 3, and 5, and Eq. (7) uses Conv 3x3,d1, Conv 3x3,d3, and Conv 3x3,d5. However, Table 1 lists a dilation value of 7 for layers 1 through 16. These two descriptions cannot both be correct. The authors should clarify the exact dilation rates used and re-derive the receptive field row accordingly.","section":"Section 4.2 and Table 1"},{"comment":"The paper states that all methods were 'performed in the same training and testing datasets for fair comparison,' but does not specify whether baseline models were retrained by the authors or evaluated using pretrained models from their original releases. Given the duplicate baseline values in Table 2, the provenance of every baseline number needs to be stated explicitly. Without this information, the fair-comparison claim and the resulting state-of-the-art claim are not fully supported.","section":"Section 5.2, Table 2"}],"minor_comments":[{"comment":"The sentence 'three two synthetic datasets' contains a typo and should read 'three synthetic datasets.'","section":"Section 5.1"},{"comment":"Minor typographical issues include 'min-batch' (should be 'mini-batch'), 'Table. 4' (should be 'Table 4'), and 'Tabel 1' in the caption of Table 1.","section":"Section 5.1 and Table 4"},{"comment":"The text says the comparison includes UGSM [29], but UGSM does not appear in Table 2; it appears only in the real-world qualitative figures. The authors should either add UGSM to the quantitative table or remove it from the list of compared methods.","section":"Section 5.2"},{"comment":"The caption of Fig. 1 is duplicated in the manuscript text; please remove the repeated caption block.","section":"Fig. 1"},{"comment":"The phrase 'reception field' is used in several places; the standard term is 'receptive field.'","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The duplicate baseline values in Table 2 and Fig. 1 are the main obstacle to acceptance. If the authors can supply corrected and independently verified baseline numbers, and if the input specification of the detail repair network is clarified, the paper could be suitable for publication. If the duplicate values turn out to be unreproducible transcription errors that materially change the reported rankings, the central claim would be undermined."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nDRD-Net is a two-stage deraining network: a rain residual network with SE blocks, then a detail repair network built from dilated-context aggregation blocks. The detail-repair idea is the actual contribution — most deraining methods just try not to lose details, this one explicitly tries to recover them after the fact. The ablation study in Table 3 supports that both the SE operation and the SDCAB help, and the code is public. That is real work, and the paper is clearly written.\n\nThe problem is the quantitative results. The stress-test note is right: Table 2 lists identical PSNR/SSIM for DDN and GMM on Rain800 (24.04 / 0.8675 for both), and Fig. 1 lists identical values for GMM and DSC (10.61 / 0.4675). Two different methods producing bit-for-bit identical scores on the same data is not credible. Something was copied wrong. And the claimed advantage over RESCAN on Rain200L is 37.15 vs 37.07 — one small correction could flip the ranking. The paper does not state whether baselines were rerun or taken from originals, and there are no error bars.\n\nThe central claim that DRD-Net outperforms state-of-the-art therefore rests on numbers that are currently unreliable. That does not mean the method is bad; the architecture is reasonable and the visual comparisons are suggestive. But \"state-of-the-art\" cannot be asserted until the baselines are independently reproduced or corrected.\n\nAlso, the novelty claim is overstated. The \"for the first time\" language about two subnetworks is mild, but SE and dilated context aggregation are exactly the ingredients in RESCAN, so the delta over that work is the detail repair module — legitimate but incremental.\n\nMinor: the real-world evaluation is purely visual; the Google Vision API experiment is nice but peripheral. The limitation section is honest about parameters, but doesn't mention the synthetic-to-real gap.\n\nBottom line: this deserves peer review, but only with the table numbers checked. I'd ask the authors to provide reproduction scripts or corrected baselines, and to soften the novelty positioning. If the duplicate values turn out to be transcription errors and the corrected baselines hold, this is a solid incremental contribution. If the margins shrink further, it becomes a minor one.\n\nRecommendation: send to peer review with a request for baseline verification.","headline":"DRD-Net's detail-repair idea is plausible and the ablation supports it, but duplicated baseline numbers make the state-of-the-art claim unreliable until corrected.","tokens_in":15430,"tokens_out":1855,"would_cite":false,"duration_ms":17269,"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 two-stage network that first removes rain streaks and then repairs the details those removals blur can outperform existing single-image deraining methods.","keywords":["single image deraining","detail recovery","squeeze-and-excitation","dilated convolution","context aggregation","residual learning","rain streak removal","image restoration"],"falsifier":"Run DRD-Net on a paired real-world rainy/clean benchmark or a physically based rain renderer that breaks O = B + R with scattering and motion-blurred streaks; if its PSNR/SSIM advantage over earlier methods disappears, or the detail-repair subnetwork outputs near zero when given a clean image (i.e., it hallucinates rather than recovers actual lost detail), the paper's central claim is falsified.","tokens_in":14466,"feed_emoji":"🌧️","tokens_out":5957,"duration_ms":59999,"temperature":0.7,"pith_summary":"This paper tries to show that the detail loss caused by single-image deraining is not an unavoidable side effect: it can be explicitly reversed by a second subnetwork dedicated to detail repair. The proposed DRD-Net first learns to predict rain streaks R from a rainy image O and subtracts them, then a detail repair network built from dilated-context blocks adds back the fine structure that the subtraction blurred. On the three synthetic benchmarks Rain200L, Rain200H, and Rain800, the full model reports higher PSNR and SSIM than six earlier methods, and the authors show examples where fine details such as zebra stripes stay intact. If correct, this reframes deraining as a two-step inverse problem rather than a single mapping, and suggests that explicit detail recovery is worth adding to other image restoration tasks.","feed_headline":"Deraining cleans rain, then a repair net restores lost detail","feed_subtitle":"A second subnetwork adds back the edges that rain removal blurs, lifting PSNR and SSIM on three benchmarks.","key_machinery":"Two building blocks carry the argument. The rain residual network uses rain residual blocks: a standard residual block wrapped in a squeeze-and-excitation (SE) operation, which learns a per-channel weight and lets the network emphasize feature channels carrying spatial context; the paper visualizes that the highest-weighted feature maps are indeed the context-rich ones. The detail repair network stacks 16 structure detail context aggregation blocks (SDCABs): each SDCAB runs three dilated convolutions with dilation rates 1, 3, and 5 on the same feature map, concatenates their outputs, compresses with a 1×1 convolution, and adds the input back. These dilated rates give the stack a receptive field of 231×231 at depth 35, which the paper argues is large enough to gather the context needed to reconstruct details rather than hallucinate them.","core_discovery":"The central claim is that a derained image has lost real, recoverable information, and that a network can be trained to predict precisely that lost information. Formally, the rainy observation is modeled as O = B + R; the rain residual network learns a function f with f(O) ≈ R, giving a preliminary derained image I_p = O - R. The detail repair network then learns a function g with I_p + g(O) ≈ B, so g(O) is the network's estimate of the detail component destroyed by removing R. The two losses are combined as λ1 ||f(O) - R||² + λ2 ||I_p + g(O) - B||² with λ1 = 0.1 and λ2 = 1.0. The authors argue that the second subnetwork, not just a deeper first network, is what lets the method preserve small high-frequency structures, and they support this with ablations showing the full two-subnetwork model beats the rain residual network alone on all three datasets.","pith_inferences":["Our inference: the two-stage remove-then-repair scheme is not specific to rain; the same SDCAB-based repair network could be attached to any restoration front end that blurs high-frequency structure, such as dehazing or denoising, and the ablation logic suggests it would help there too.","Our inference: the paper's difference image Y-X implies the detail network acts as a high-frequency corrector, so one testable prediction is that the energy of g(O) concentrates along edges, which could be verified by gradient-magnitude statistics.","Our inference: because the architecture depends on the additive model O = B + R, the method's real-world advantage is not guaranteed by these experiments; a fair test would use paired real rain/clean images or a physically based renderer with scattering and motion-blurred streaks.","Our inference: if real rain follows the atmospheric-scattering variant O = α(B + ΣR_i) + (1−α)A, the detail repair network may need the transmission α as an additional input to correctly recover details in foggy heavy-rain scenes."],"forward_implications":["If the central claim holds, deraining systems should be built as two stages rather than one, with the second stage explicitly trained on the residual between the derained image and the ground truth.","The reported gains are largest on the heavy-rain set Rain200H, so the detail-repair mechanism matters most when rain and background structure overlap in the high-frequency band.","The ablation results imply that each added component—SE weighting, a residual detail network, and the dilated-context block—pays off monotonically on all three benchmarks.","The paper's downstream experiment suggests that removing rain this way lowers a commercial vision recognizer's confidence that an image contains rain, which is a direct practical benefit if confirmed."],"supporting_citations":[{"why":"Traditional GMM layer-prior baseline that DRD-Net must beat and whose layer-decomposition idea the paper extends.","marker":"[3]"},{"why":"First deep deraining baseline, also the source of the residual block architecture used in the ablation comparison.","marker":"[7]"},{"why":"Learning-based deraining baseline in all comparisons and the second-best method on Rain800.","marker":"[8]"},{"why":"Main learning-based competitor, and the source of both the SE-context idea and the dilated-context receptive field motivation.","marker":"[10]"},{"why":"Supplies the Rain200L/Rain200H synthetic datasets, real rain images, and the two rain models including the atmospheric-scattering variant.","marker":"[16]"},{"why":"Supplies the Rain800 dataset used for training and testing.","marker":"[17]"},{"why":"Defines the squeeze-and-excitation operation that the rain residual block is built on.","marker":"[21]"},{"why":"Introduces dilated-convolution context aggregation, the principle behind the SDCAB's large receptive field.","marker":"[25]"}],"fun_headline_variants":["Two-stage deraining: remove rain, then restore lost detail","DRD-Net: rain removal plus a detail-repair subnetwork","After rain removal, a repair network brings back fine details","Deraining without detail loss: a two-subnet approach","Rain removal that recovers what it blurs: DRD-Net"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes real rain is well approximated by the synthetic additive model O = B + R, with rain streaks separable from background details, so that subtracting the predicted R from O leaves a detail-lost but recoverable image; if real rain violates this assumption, the two subnetworks may simply be fitting synthetic statistics.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage deraining: remove rain, then restore lost detail","DRD-Net: rain removal plus a detail-repair subnetwork","After rain removal, a repair network brings back fine details","Deraining without detail loss: a two-subnet approach","Rain removal that recovers what it blurs: DRD-Net"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0004,"raw_usage":{"total_tokens":2130,"prompt_tokens":1029,"completion_tokens":1101,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":1013}},"tokens_in":645,"tokens_out":1101,"duration_ms":10132,"temperature":1.0,"reasoning_tokens":1013,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:47:29.544447+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DRD-Net on a paired real-world rainy/clean benchmark or a physically based rain renderer that breaks O = B + R with scattering and motion-blurred streaks; if its PSNR/SSIM advantage over earlier methods disappears, or the detail-repair subnetwork outputs near zero when given a clean image (i.e., it hallucinates rather than recovers actual lost detail), the paper's central claim is falsified.","supporting_citations":[{"cited_title":"Squeeze-and-excitation networks,","cited_arxiv_id":null,"evidence_quote":"Defines the squeeze-and-excitation operation that the rain residual block is built on."},{"cited_title":"Recurrent squeeze-and- excitation context aggregation net for single image deraining,","cited_arxiv_id":null,"evidence_quote":"Main learning-based competitor, and the source of both the SE-context idea and the dilated-context receptive field motivation."},{"cited_title":"Rain streak removal using layer priors,","cited_arxiv_id":null,"evidence_quote":"Traditional GMM layer-prior baseline that DRD-Net must beat and whose layer-decomposition idea the paper extends."},{"cited_title":"Depth-attentional features for single-image rain removal,","cited_arxiv_id":null,"evidence_quote":"Learning-based deraining baseline in all comparisons and the second-best method on Rain800."}],"review_version":1}