CMAWRNet: Multiple Adverse Weather Removal via a Unified Quaternion Neural Architecture

Karen Panetta; Peter Huang; Sos Agaian; Vladimir Frants

REVIEW 4 major objections 5 minor 3 references

CMAWRNet: Multiple Adverse Weather Removal via a Unified Quaternion Neural Architecture

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read One network removes mixed weather by splitting each image into a structure layer and a texture layer first.

desk verdict 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. read the letter →

arxiv 2505.01882 v1 pith:BUARL2PG submitted 2025-05-03 cs.CV

Vladimir Frants , Sos Agaian , Karen Panetta , Peter Huang This is my paper · ORCID

classification cs.CV

keywords adverseweatherremovaluniversalimagerestorationquaternionneuralnetworkstexture-structuredecompositionRetinexrainstreaksnowlow-lightcorrection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

The 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.

The 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.

All 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.

Referee Report

4 major / 5 minor

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.

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 (4)

[IV-B, Eqs. (5)-(6)] 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.
[IV-A and IV-B, training and validation of the decomposition] 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.
[IV-B, Tables I-IV] 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.
[Introduction and IV-C (architecture description)] 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.

minor comments (5)

[General] 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.
[IV-B, Tables III-IV] 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.
[Table V] 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.
[Eq. (14)] 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].
[References] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the decomposition identity is a stated model assumption, and all reported gains are measured against external benchmarks with held-out synthetic and real-world data.

full rationale

The paper's central derivation chain is a conventional supervised restoration setup: it defines a Retinex-style multiplicative model I = S ∘ T (Eq. 4), constructs initial maps S* and T* with T* := I ⊘ S*, and then trains DNet, TNet, and FNet on paired degraded/clean images from RainDrop, Snow100K, and Outdoor-Rain. The identity I = S* ∘ T* is indeed true by construction because T* is defined as the quotient, but the paper does not present that identity as a fitted prediction or as evidence for its weather-removal results; it is a modeling choice. The reported PSNR/SSIM numbers and the DAWN detection mAP are measured against external clean ground truth or real-world annotations, not against quantities used to define the network's parameters. The QSSIM training loss is not identical to the Y-channel SSIM or PSNR used in evaluation, and the detection experiment uses real DAWN images not used in training. The authors' self-citations ([69], [70]) appear only as comparison baselines in Table IV and in complexity analysis; they are not used to justify the architecture's uniqueness or to supply a load-bearing theorem. Missing Section III, reversed table references, and the questionable interpretation of Eq. 6 as a smooth structure map are correctness or presentation risks, not circularity. Accordingly, no step in the claimed derivation reduces to its own input by construction or by self-citation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a multiplicative image-formation model that is invoked but not independently validated. The only explicitly hand-set design parameters are gamma1 and gamma2 for the gradient-based decomposition. No new physical entities are introduced; texture and structure are internal latent modeling constructs, not independently evidenced entities.

free parameters (2)

gamma_1 = 0.5
Hand-set gradient exponent in the guidance map computation, Eq. 5, Section IV-B; controls sensitivity to texture details.
gamma_2 = 1.5
Hand-set gradient exponent in the initial structure map computation, Eq. 6, Section IV-B; modifies the impact of gradient magnitude in the structure estimate.

assumptions (4)

domain assumption The input image is multiplicatively decomposable as I = S∘T into structure and texture components.
Invoked in Section IV-A, Eq. 4, to justify processing the two branches separately; no independent validation is provided.
domain assumption Haze affects mostly the structure component, while rain streaks, raindrops, and snowflakes affect mostly the texture component.
Section IV-A uses this assignment to motivate DNet and the separate TNet branches; it is a modeling choice not measured directly.
domain assumption Models trained on synthetic weather datasets (Outdoor-Rain, Snow100K, RainDrop) generalize to real-world weather images such as DAWN.
Section IV-A describes training on synthetic data; Section IV-B evaluates on DAWN and real photographs without domain adaptation.
domain assumption The atmospheric scattering relation J(x)=M(x)I(x), with M estimated from airlight and transmission, is a valid model for the structure branch.
Eqs. 10 and 11 in Section IV-C follow standard dehazing models but are assumed as the forward model for the latent variable M.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CMAWRNet: Multiple Adverse Weather Removal via a Unified Quaternion Neural Architecture." pith.science (2026). https://pith.science/paper/BUARL2PG

@misc{pith2026250501882,
  author       = {Pith},
  title        = {Pith review of: CMAWRNet: Multiple Adverse Weather Removal via a Unified Quaternion Neural Architecture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BUARL2PG}},
  note         = {Machine review of arXiv:2505.01882}
}

read the original abstract

Images used in real-world applications such as image or video retrieval, outdoor surveillance, and autonomous driving suffer from poor weather conditions. When designing robust computer vision systems, removing adverse weather such as haze, rain, and snow is a significant problem. Recently, deep-learning methods offered a solution for a single type of degradation. Current state-of-the-art universal methods struggle with combinations of degradations, such as haze and rain-streak. Few algorithms have been developed that perform well when presented with images containing multiple adverse weather conditions. This work focuses on developing an efficient solution for multiple adverse weather removal using a unified quaternion neural architecture called CMAWRNet. It is based on a novel texture-structure decomposition block, a novel lightweight encoder-decoder quaternion transformer architecture, and an attentive fusion block with low-light correction. We also introduce a quaternion similarity loss function to preserve color information better. The quantitative and qualitative evaluation of the current state-of-the-art benchmarking datasets and real-world images shows the performance advantages of the proposed CMAWRNet compared to other state-of-the-art weather removal approaches dealing with multiple weather artifacts. Extensive computer simulations validate that CMAWRNet improves the performance of downstream applications such as object detection. This is the first time the decomposition approach has been applied to the universal weather removal task.

Figures

Figures reproduced from arXiv: 2505.01882 by the authors.

**Figure 1.** Comparison between RNN and QNN for image restoration. a) Input image with degradation. b) Output from the RNN (0.3M parameters). c) Ground truth. d) Output from the QNN (0.078M parameters) shows improved color fidelity and reduced geometric artifacts while significantly lowering the number of parameters. II. PREVIOUS WORK A. Adverse Weather Removal Existing image restoration approaches include single weather removal… view at source ↗

**Figure 2.** The proposed framework for CMAWRNet. The enhancement process is divided into decomposition, transformation, and fusion. In the decomposition step, a subnetwork DNet decomposes the input image into texture and structure—two separate instances of encoder-decoder based TNet clean texture and structure images in the following cleaning step. Feature sharing between instances of TNet is introduced to ensure adequate proce… view at source ↗

**Figure 3.** Comparison of synthetic rain and haze removal by CMAWRNet and state-of-the-art methods. Specialized haze removal techniques, EPDN and RefineDNet, effectively remove haze from most images but leave rain streaks and produce darker images with loss of detail in dark regions, especially struggling with non-homogeneous haze structures like in image d). MPRNet removes rain streaks but yields low-contrast images with resid… view at source ↗

Figures from the paper (3 more)

**Figure 5.** Figure 5: Rain drop images. a) Input, b) – TransWeather, c) CMAWRNet, d) Ground Truth. Both TransWeather and CMAWRNet successfully remove raindrops, but CMAWRNet produces a brighter image with better details. TABLE I QUANTITATIVE COMPARISON ON THE TEST1 (RAIN+FOG) DATASET BASED …

**Figure 6.** Figure 6: Detection results on the image without preprocessing (Baseline), Transweather, and CMAWRNet. As can be seen, CMAWRNet improves the detection of small background objects in fog, snow, and rain conditions. As can be seen, CMAWRNet improves the detection performance in al…

**Figure 7.** Figure 7: a) input image, b) Chen et. al. c) TransWeather, d) CMAWRNet. CMAWRNet generally produces vivid images with better visibility. CMAWRNet brings up better background details in the image (1). Moreover, in image (2) both TransWeather and Chen et al. methods failed to remo…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

[3]

Towards Real-time High-Definition Image Snow Removal: Efficient Pyramid Network with Asymmetrical Encoder-decoder Architecture

J. Zhang et al., “HazDesNet: An end-to-end network for haze density prediction,” IEEE Trans. Intell. Transp. Syst., vol. 23, no. 4, pp. 3087-3102, 2020. [4] B. Li et al., “An all-in-one network for dehazing and beyond,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), 2017, pp. 4770–4778. [5] S. Zhao et al., “RefineDNet: A weakly supervised refinement framew...

work page Pith review arXiv 2022
[29]

A survey of quaternion neural networks,

T. Parcollet, M. Morchid, and G. Linares, “A survey of quaternion neural networks,” Artif. Intell. Rev., 2019. [30] A. Grigoryan and S. Agaian, Quaternion and Octonion Color Image Processing With MATLAB, SPIE Press Monographs, 2018. [31] A. Grigoryan and S. Agaian, “Optimal restoration of multiple signals in quaternion algebra,” in Proc. SPIE Mob. Multime...

work page 2019
[56]

Learning to Remove Bad Weather: Towards Robust Visual Perception for Self-Driving,

Y. Lee, et al., “Learning to Remove Bad Weather: Towards Robust Visual Perception for Self-Driving,” IEEE Robot. Autom. Lett., 2022. [57] R. Li, et al., “Heavy rain image restoration: Integrating physics model and conditional adversarial learning,” in Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 1633–1642. [58] C. Wei, et al., “D...

arXiv 2022

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [3]

Towards Real-time High-Definition Image Snow Removal: Efficient Pyramid Network with Asymmetrical Encoder-decoder Architecture

J. Zhang et al., “HazDesNet: An end-to-end network for haze density prediction,” IEEE Trans. Intell. Transp. Syst., vol. 23, no. 4, pp. 3087-3102, 2020. [4] B. Li et al., “An all-in-one network for dehazing and beyond,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), 2017, pp. 4770–4778. [5] S. Zhao et al., “RefineDNet: A weakly supervised refinement framew...

work page Pith review arXiv 2022

[2] [29]

A survey of quaternion neural networks,

T. Parcollet, M. Morchid, and G. Linares, “A survey of quaternion neural networks,” Artif. Intell. Rev., 2019. [30] A. Grigoryan and S. Agaian, Quaternion and Octonion Color Image Processing With MATLAB, SPIE Press Monographs, 2018. [31] A. Grigoryan and S. Agaian, “Optimal restoration of multiple signals in quaternion algebra,” in Proc. SPIE Mob. Multime...

work page 2019

[3] [56]

Learning to Remove Bad Weather: Towards Robust Visual Perception for Self-Driving,

Y. Lee, et al., “Learning to Remove Bad Weather: Towards Robust Visual Perception for Self-Driving,” IEEE Robot. Autom. Lett., 2022. [57] R. Li, et al., “Heavy rain image restoration: Integrating physics model and conditional adversarial learning,” in Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 1633–1642. [58] C. Wei, et al., “D...

arXiv 2022