{"id":"8ea73f74-5be7-4363-a31e-e960ae0082f0","arxiv_id":"2505.01882","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CMAWRNet, a unified quaternion neural architecture with texture-structure decomposition and attentive fusion, reports better PSNR/SSIM than prior universal weather-removal models on rain-fog, snow, and raindrop benchmarks.","lead":"A new image-restoration system, CMAWRNet, is designed to clear haze, rain streaks, snow, and raindrops in a single pass using a quaternion-based neural network and a texture/structure decomposition. The authors report that it beats current all-in-one weather-removal networks on standard benchmarks and improves object detection on real traffic images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 6 defines S* as a spatially averaged gradient-magnitude map, not a smooth structure component; dividing I by S* cannot yield a meaningful texture component, so the central decomposition of Eq. 4 is internally inconsistent.","rationale":"The reader's weakest assumption correctly identifies the decomposition as the load-bearing element of the method. My reading sharpens this: the problem is not only whether real images satisfy I = S∘T, but whether the paper's own initial construction implements such a factorization. Eq. 6 produces a gradient-magnitude average, which cannot serve as a smooth structure image for arbitrary color images; the subsequent division is numerically unstable in flat regions. This is an internal consistency issue rather than a dispute with consensus, and it directly affects the architecture's claimed separation of weather degradations into structure vs texture channels. I still keep the verdict CONDITIONAL because the empirical tables may survive even if the proposed decomposition is misdescribed: the end-to-end training could in principle let DNet learn a useful split from the LOL/weather data despite a poor initialization. However, as written, the central mechanism is not reproducible or verifiable without code, so the paper should not be accepted without a corrected derivation or artifact. The concrete test would settle whether Eq. 6 is a typographical/OCR artifact (in which case the concern dissolves) or a genuine flaw (in which case the method's justification must be rewritten). I did not base the objection on the missing Section III, reversed table labels, or duplicated text, though those issues compound the verifiability problem.","tokens_in":15106,"tokens_out":7182,"duration_ms":78442,"concrete_test":"Implement Eqs. 5-6 exactly as written on a sample image from the Outdoor-Rain Test1 set that contains a flat sky region: compute S* with γ1=0.5, γ2=1.5 and T* = I ⊘ S*. Verify (a) S* is a smooth RGB structure image, (b) T* is finite and texture-like with values in a plausible range, and (c) S*∘T* equals the input I. If S* is instead a gradient-magnitude map that collapses to zero in flat areas, the factorization as published is invalid and the two-branch cleaning premise fails; the authors should then provide the corrected decomposition or code demonstrating the intended formula.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's two-branch design is justified by Eq. 4, I = S∘T, where S is said to be the smooth structure/illumination component and T the texture component, with haze acting on S and rain/snow/raindrops on T. But the prescribed initialization in Eqs. 5-6 computes S* as (1/|Ω|) Σ_{Ω} |∇I|^{γ2}, i.e., a local average of an exponentiated gradient magnitude. That is an edge-strength map, not a smooth image component. In flat regions S* tends to zero, so T* = I ⊘ S* is undefined or noise-amplifying; in textured regions S* is large, so T* carries no consistent photometric relation to the original image. The claimed multiplicative reconstruction S*∘T* = I cannot be maintained as a meaningful structure/texture split. Because DNet receives these initial maps and is trained on low-light LOL pairs without structure/texture ground truth, there is no evidence the learned decomposition separates the physical degradations the paper attributes to S and T. If the equations are a typo, the paper must supply the correct decomposition; if they are not, the core mechanism is unsupported. The missing Section III, reversed table references, and absent code make this impossible to resolve from the text alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CMAWRNet, a unified quaternion neural network for removing multiple adverse weather degradations (haze combined with rain streaks, snow, and adherent raindrops) in a single forward pass. The architecture consists of DNet, which decomposes an input image into structure and texture components; two quaternion transformer encoder-decoder instances (TNet-H and TNet-S), which clean the two components with feature sharing; FNet, which performs attentive fusion with low-light correction; and a quaternion structural similarity (QSSIM) training loss intended to preserve color information. The paper reports PSNR/SSIM improvements over TransWeather and Chen et al. on the Outdoor-Rain Test1, Snow100K-L, and RainDrop benchmarks, an ablation study showing the contribution of each module, and an object-detection experiment on the real-world DAWN dataset indicating improved mAP after restoration.","tokens_in":15303,"tokens_out":5912,"duration_ms":57405,"significance":"If the reported results hold, this paper would make a moderate but useful contribution: a single lightweight network that outperforms two established universal weather-removal baselines on three standard benchmarks, with a color-aware quaternion loss and downstream detection validation on real-world images. The strengths include the use of real DAWN images for the detection experiment, the explicit quaternion formulation, and a component-level ablation. However, the central decomposition mechanism is under-specified and appears internally inconsistent (Eq. 6), the experiments are reported as point estimates without code, error bars, or significance tests, and the manuscript has structural errors such as a missing Section III and reversed table references. These issues currently prevent verification of the central performance claim.","major_comments":[{"comment":"Section IV-B defines the initial structure map S* as the local average of exponentiated gradient magnitudes, S* = (1/|Ω|)Σ_{Ω}|∇I|^{γ2}. This quantity is an edge-strength map, not a smooth structure or illumination component as required by the multiplicative model I = S∘T. Consequently, T* = I ⊘ S* is not a meaningful texture component: in flat regions S* tends to zero, making the division unstable or noise-amplifying, while in textured regions T* does not maintain a consistent photometric relation to the original image. Because the two-branch cleaning architecture is motivated entirely by this decomposition, the core mechanism of the paper is unsupported unless the correct form of Eq. (6) is supplied and the decomposition is validated, for example by visualizing S and T on real images and checking that I ≈ S∘T holds after DNet.","section":"IV-B, Eqs. (5)-(6)"},{"comment":"The training protocol described in Section IV-A is not sufficient to establish that DNet learns the intended structure/texture separation. DNet and FNet are first trained on LOL low-light pairs, but neither LOL nor the weather datasets provide ground-truth structure/texture decompositions, and no loss term or auxiliary constraint is specified to enforce that S is smooth and T is texture-only. The claim that haze corrupts S while rain, snow, and raindrops corrupt T is therefore an assumption that is validated only indirectly through end-to-end PSNR/SSIM. The reported gains could in principle be attributed to the synthetic training distribution rather than to the decomposition. The authors should provide visual decomposition results, a quantitative check of the multiplicative reconstruction, and ideally an evaluation on real images with known degradation sources.","section":"IV-A and IV-B, training and validation of the decomposition"},{"comment":"The experimental results are reported as single point estimates of PSNR/SSIM and mAP/mAR with no error bars, no number of runs, and no significance tests, and no code is released. Several reported improvements are modest (e.g., RainDrop SSIM 0.9518 vs. 0.9460 for TransWeather in Table III; DAWN rain mAP 0.536 vs. 0.520 baseline in Table IV), so without statistical evidence and implementation details the claimed advantages cannot be verified. In addition, the table references are internally inconsistent: the RainDrop text refers to 'Table IV' but the raindrop results are in Table III, and the object-detection text refers to 'Table III' but the DAWN results are in Table IV. These errors must be corrected and the numerical results must be reproducible before the central performance claim can be accepted.","section":"IV-B, Tables I-IV"},{"comment":"The paper is missing its architecture section: the introduction states that Section III describes the proposed method, but the manuscript jumps from Section II to Section IV, and the algorithm referred to as 'Algorithm 1' in Section IV-A is never included. Key architectural details are also underspecified: the exact configuration of the quaternion transformer blocks (number of heads, stages, hidden dimensions), the feature-sharing mechanism between TNet-H and TNet-S, the channel dimensions of the quaternion convolutions, and the role of Eqs. (10)-(11) with the latent variable M are not described with enough precision to allow independent implementation. This lack of detail prevents replication and evaluation of the method.","section":"Introduction and IV-C (architecture description)"}],"minor_comments":[{"comment":"Section numbering is inconsistent: there is no Section III despite multiple references to it, and the Experimental Results section is numbered IV while the Conclusion is numbered V; the sections should be renumbered.","section":"General"},{"comment":"The RainDrop results are called 'Table IV' in the text but appear in Table III, and the DAWN object-detection results are called 'Table III' in the text but appear in Table IV; these cross-references should be corrected.","section":"IV-B, Tables III-IV"},{"comment":"The caption of Table V reads 'AVERAGE MEAN PRECISION AND RECALL ON SUBSETS OF DAWN DATASET' but the table contains ablation results with PSNR, SSIM, parameters, and inference time; the caption should be revised to match the content.","section":"Table V"},{"comment":"Equation (14) uses notation such as quaternion means, variations, and covariance without defining how these are computed from quaternion image components; please either define these quantities or cite the specific equations from [57].","section":"Eq. (14)"},{"comment":"Reference [66] lists the DAWN dataset with an arXiv link to paper 2207.05605, which is not the DAWN paper; the correct arXiv identifier appears to be 2008.05402.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This manuscript reads as an early draft: it lacks Section III, mislabels several tables, and contains an incorrect caption for the ablation table. The competitive claims are plausible but not verifiable without code and multi-run statistics. The decomposition inconsistency in Eq. (6) is the main technical concern; if the authors can supply a corrected formulation and reproducible experiments, a revision could be publishable, but in its current form the central mechanism is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, two things up front. First, the paper reports real-looking gains on synthetic weather benchmarks and a clean ablation story, and the quaternion angle is a reasonable thing to try for multi-weather removal. Second, the central decomposition is internally inconsistent. Eq. 4 says I = S∘T, with S the smooth structure and T the texture, but Eq. 6 defines S* as a local average of exponentiated gradient magnitude. That is an edge-strength map, not a smooth image component. In flat regions S* tends to zero, so T* = I ⊘ S* is undefined or noise-amplifying; in textured regions T* carries no consistent photometric meaning. The claim that the network separates the physical degradations is unsupported by this math. If the equations are a typo, the authors need to supply the correct decomposition; if not, the two-branch design has no stated justification.\n\nThe paper does have strengths. The architecture is well-specified: DNet, TNet with quaternion transformers and shared features, attentive FNet with low-light correction, and a QSSIM loss. The ablation in Table V shows each component adds something. On Outdoor-Rain Test1, Snow100K-L, and RainDrop the reported PSNR/SSIM beat TransWeather and Chen et al. by nontrivial margins, and the DAWN detection numbers point the right way, especially for small objects. That is enough to take the method seriously.\n\nThe soft spots are proportionate to that. No code, no checkpoints, no error bars, no statistical significance tests, just single-point tables. The text has at least two internal errors: Table III and Table IV references are swapped, and there is no Section III despite the paragraph promising one. A paragraph about visual results is duplicated. More importantly, DNet is trained on LOL low-light pairs without any structure/texture ground truth, so there is no evidence the learned decomposition matches the physics the paper assigns to S and T. The detection evaluation uses SCNet but the description of how categories are filtered is thin.\n\nAll of this is addressable. The paper is a candidate for peer review, not a desk reject, but a referee should insist on a corrected decomposition definition, a released artifact or at least detailed training/inference code, and a cleanup of the references and duplicated text. If the mechanism can be fixed, this would be a useful contribution to weather removal and quaternion networks; as it stands, the central claim rests on a formula that does not do what the paper says it does.","headline":"The quaternion weather-removal architecture is plausible and the reported numbers look good, but the core decomposition equation is self-contradictory and the paper lacks code, error bars, and clean exposition; as submitted it needs major revision before the central claims can be trusted.","tokens_in":15863,"tokens_out":4009,"would_cite":false,"duration_ms":37791,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One network removes mixed weather by splitting each image into a structure layer and a texture layer first.","keywords":["adverse weather removal","universal image restoration","quaternion neural networks","texture-structure decomposition","Retinex decomposition","rain streak removal","snow removal","low-light correction"],"falsifier":"Take a real mixed-weather photograph with a smooth, textureless region (a fogged wall or sky) that also contains visible rain streaks, and inspect DNet's outputs: if the structure map $S$ retains rain-streak edges or the texture map $T$ retains large-scale haze gradients, the separation that motivates the two branches is violated. A quantitative version: build a held-out set of such images, compare CMAWRNet against a single-branch quaternion transformer with the same parameter count, and see whether the decomposition-based model still wins on PSNR/SSIM; if it does not, the decomposition, not the quaternion machinery, is the suspected source of the reported gains.","tokens_in":14856,"feed_emoji":"🌦️","tokens_out":12263,"duration_ms":113273,"temperature":0.7,"pith_summary":"CMAWRNet claims that universal weather-removal networks fail on mixed weather because they try to undo all degradations in one entangled representation. Its fix is to decompose every input image multiplicatively, $I = S \\circ T$, into a smooth structure layer $S$ and a texture layer $T$, on the Retinex-style assumption that haze mostly corrupts $S$ while rain streaks, snow, and adherent raindrops mostly corrupt $T$. Two lightweight quaternion transformer branches clean the two layers separately with shared features, and an attentive quaternion fusion block recombines them while correcting low light. The paper reports higher PSNR/SSIM than prior universal baselines on the rain-plus-haze, snow, and raindrop test sets, plus improved small-object detection on real traffic images from the DAWN dataset. If this holds, one fixed set of weights can replace the separate dehazing, deraining, and desnowing modules that current systems run.","feed_headline":"One network clears haze, rain, snow, and raindrops","feed_subtitle":"A structure-texture split lets one model beat prior universal weather removers and sharpen small-object detection.","key_machinery":"The load-bearing object is the Retinex-style multiplicative decomposition $I = S \\circ T$ (Eq. 4), computed by DNet from exponentiated local gradients with hand-set exponents $\\gamma_1 = 0.5$ and $\\gamma_2 = 1.5$. This identity carries the whole argument because it is what lets the architecture assign haze to the structure branch and rain, snow, and raindrops to the texture branch. Around that split, the design uses quaternion algebra, treating RGB as one quaternion number and using Hamilton-product convolutions, to preserve inter-channel color relations; a shared-feature quaternion transformer (TNet) with one instance per layer; attention masks in FNet that weight structure versus texture during fusion; and a quaternion structural-similarity loss that scores color as a whole rather than channel by channel.","core_discovery":"The paper's central claim is that a single network can handle several co-occurring weather degradations by separating the image into what it calls structure and texture before any cleaning happens. With the multiplicative model $I = S \\circ T$, the structure map $S$ captures the scene's smooth, large-scale layout and the texture map $T$ captures fine detail; the authors assert that haze acts mainly on $S$ and that rain, snow, and raindrops act mainly on $T$. Two quaternion transformer encoder-decoders, one per layer and sharing features, produce cleaned versions of each layer, and the attentive fusion network FNet weights and recombines them while applying gamma correction. The paper's measured result is a single 18-million-parameter model that reaches 30.02 dB PSNR on the rain-plus-haze Test1 set, 30.08 dB on Snow100K-L, and 32.43 dB on RainDrop, each above the compared universal baselines, and that raises mean average precision on the DAWN weather image set, most clearly for small objects.","pith_inferences":["The paper leaves the two decomposition exponents $\\gamma_1$ and $\\gamma_2$ fixed by hand; a natural extension is to learn them per image or per weather type, since the whole two-branch argument depends on the initial split being right.","If the structure-versus-texture split is genuinely what drives the gain, the same decomposition could be tested on other mixed degradations that separate by scale, such as blur plus sensor noise, even though the paper only demonstrates weather.","The downstream evaluation is limited to object-detection mAP at IoU 0.5 with the SCNet detector; a reader should not assume the same improvement transfers to semantic segmentation or detection at stricter IoU thresholds.","The paper argues that standard SSIM ignores color and uses a quaternion loss, yet the headline benchmark tables still report PSNR and SSIM on the Y channel; checking the restored images with a color-aware metric would test whether the quaternion loss actually buys what the authors claim."],"forward_implications":["If the benchmark numbers hold, a single trained CMAWRNet removes the difficult combination of dense haze plus rain streaks in one forward pass, a case the paper says prior universal models handle poorly.","The same fixed weights reach the top PSNR/SSIM on the snow test set and the raindrop test set, so no condition-specific encoder or fine-tuning is needed for those three weather types.","Restoring images before detection with CMAWRNet improves mAP and mAR on the DAWN traffic images, with the largest gains in small-object metrics (mAPS, mARS), which matters for surveillance and autonomous driving.","The reported 24.3 ms inference time on a 1920x1080 frame is lower than the compared universal baselines, so the accuracy improvement is not bought with extra runtime.","The ablation study attributes the gain to every component, including decomposition, shared features, quaternion layers, and the quaternion similarity loss, so the paper's case for the full architecture depends on all four."],"supporting_citations":[{"why":"Supplies the quaternion algebra with the Hamilton product and the quaternion color representation that all quaternion layers in the architecture assume.","marker":"[30]"},{"why":"Supplies the Test1 (Outdoor-Rain) rain-plus-haze evaluation set used for the headline mixed-weather comparison.","marker":"[57]"},{"why":"Supplies the LOL low-light image pairs used to pre-train the decomposition and fusion subnetworks.","marker":"[58]"},{"why":"The visual-transformer universal baseline whose PSNR/SSIM marks CMAWRNet must beat across all three benchmark conditions.","marker":"[22]"},{"why":"The two-stage knowledge-transfer universal baseline that defines the prior state for mixed-degradation removal.","marker":"[23]"},{"why":"Supplies the Snow100K training images and Snow100K-L test set used for the snow-removal comparison.","marker":"[13]"},{"why":"Supplies the RainDrop dataset and the attentive-GAN baseline for adherent-raindrop removal.","marker":"[9]"},{"why":"The structure-and-texture-aware Retinex model that motivates the $I = S \\circ T$ decomposition used by DNet.","marker":"[63]"},{"why":"The COCO-trained detector used to measure whether CMAWRNet restorations improve object-detection mAP/mAR on real weather images.","marker":"[65]"},{"why":"Supplies the DAWN adverse-weather traffic image set and its detection annotations used in the downstream evaluation.","marker":"[66]"}],"fun_headline_variants":["One quaternion net clears haze, rain, snow","Structure-texture split makes one net beat all weather","Single network, quaternion attention: all weather gone","CMAWRNet: one model, multiple weather removal","Haze, rain, snow? One network clears them all"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole two-branch design collapses if a real bad-weather image cannot be factored into a smooth structure layer and a texture layer, with haze living only in structure and rain, snow, and raindrops living only in texture; the initial split itself is seeded by hand-chosen exponents $\\gamma_1=0.5$ and $\\gamma_2=1.5$, so the decomposition is only as good as that assumption.","fun_headline_variants_meta":{"raw":{"variants":["One quaternion net clears haze, rain, snow","Structure-texture split makes one net beat all weather","Single network, quaternion attention: all weather gone","CMAWRNet: one model, multiple weather removal","Haze, rain, snow? One network clears them all"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000274,"raw_usage":{"total_tokens":1664,"prompt_tokens":993,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":593}},"tokens_in":609,"tokens_out":671,"duration_ms":6027,"temperature":1.0,"reasoning_tokens":593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:08:03.988045+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real mixed-weather photograph with a smooth, textureless region (a fogged wall or sky) that also contains visible rain streaks, and inspect DNet's outputs: if the structure map $S$ retains rain-streak edges or the texture map $T$ retains large-scale haze gradients, the separation that motivates the two branches is violated. A quantitative version: build a held-out set of such images, compare CMAWRNet against a single-branch quaternion transformer with the same parameter count, and see whether the decomposition-based model still wins on PSNR/SSIM; if it does not, the decomposition, not the quaternion machinery, is the suspected source of the reported gains.","supporting_citations":[],"review_version":1}