Pith. sign in

REVIEW 4 major objections 4 minor 23 references

QRetinex-Net: Quaternion-Valued Retinex Decomposition for Low-Level Computer Vision Applications

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims to introduce the first quaternion-valued Retinex decomposition, rewriting the classic split of an image into reflectance and illumination as a product of two quaternion factors, $S=R\otimes I$, where $\otimes$ is the…

desk verdict Interesting quaternion Retinex idea, but the core decomposition is not actually verified because the math allows a trivial solution and the evaluation is circular. read the letter →

arxiv 2507.16683 v1 pith:CWCXGGXO submitted 2025-07-22 cs.CV

Sos Agaian , Vladimir Frants This is my paper
classification cs.CV
keywords quaternionRetinexdecompositionlow-lightimageenhancementreflectanceconsistencyindexHamiltonproductwaveletdomaincolorconstancyinfrared-visiblefusion
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

This paper claims to introduce the first quaternion-valued Retinex decomposition, rewriting the classic split of an image into reflectance and illumination as a product of two quaternion factors, $S=R\otimes I$, where $\otimes$ is the Hamilton product, computed in the Haar wavelet domain. The claim is that this single algebraic change fixes four weaknesses of scalar Retinex: processing color channels independently, lacking a model of human color perception, failing to reconstruct the input exactly, and not explaining color constancy. The payoff, if true, is a low-light image decomposition that both reproduces the input almost perfectly and keeps reflectance stable as illumination changes, which in turn improves crack segmentation, face detection, and infrared-visible fusion by 2-11 percent over prior Retinex networks.

What carries the argument

The load-bearing object is the Hamilton product of two quaternions, defined so that $q_1\otimes q_2$ mixes the scalar and three imaginary parts of each factor. Each pixel becomes a pure imaginary quaternion $0+R\mathbf{i}+G\mathbf{j}+B\mathbf{k}$; the initial reflectance map stores normalized color ratios $R/M$, $G/M$, $B/M$ with $M=\max(R,G,B)$, and the initial illumination map stores the original RGB magnitudes. A Haar wavelet transform moves both maps into a 32-channel multiscale representation, a symmetric cross-attention block lets the reflectance and illumination branches exchange information, and a multi-term loss (reconstruction, mutual reconstruction, illumination smoothness, reflectance consistency, and frequency regularization) drives the factorization. The RCI metric linearly interpolates between low- and normal-light images, decomposes each intermediate, takes the pixel-wise variance of reflectance across the interpolation, and normalizes the worst-case variance by the theoretical maximum 0.25.

What would settle it

Compute the Hamilton product of the two initial maps in equations (4) and (6): both are purely imaginary and proportional, so their product has zero imaginary part and cannot reconstruct the RGB image, meaning the learned refinement carries the entire decomposition. A decisive test is to train the same architecture with the reflectance-consistency and mutual-reconstruction losses removed: if the reconstruction PSNR stays near 62 dB while RCI drops well below 0.988, the Hamilton product alone is not producing the claimed reflectance stability.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that an RGB image can be represented as a purely imaginary quaternion $S=0+R\mathbf{i}+G\mathbf{j}+B\mathbf{k}$ and factorized as $S=Q_1\otimes Q_2$, with $Q_1$ acting as reflectance and $Q_2$ as illumination, so that the Hamilton product binds the three color channels together instead of treating them separately. Working in the Haar wavelet domain with a learned refinement network, the authors report near-perfect reconstruction on LOLv1 (PSNR 62.515 dB, SSIM 0.9998, LPIPS 0.0001) and introduce the Reflectance Consistency Index (RCI), which reaches 0.988, as evidence that their reflectance map stays almost unchanged when illumination is interpolated from dark to normal. They further report that swapping their decomposition into existing pipelines improves low-light crack segmentation, zero-shot day-night face detection, and infrared-visible fusion, outperforming RetinexNet, KIND++, URetinexNet, and Diff-Retinex by 2-11 percent across tasks.

Load-bearing premise

The network is assumed to learn non-trivial quaternion factors that simultaneously reconstruct the image and separate reflectance from illumination, even though the reconstruction objective alone would also be perfectly satisfied by the trivial split where one factor equals the input image and the other equals the number one.

Editorial extensions

If this is right

  • The same Hamilton-product factorization yields near-perfect reconstruction (62.5 dB PSNR and 0.9998 SSIM on LOLv1), so downstream tasks can use reflectance without paying a reconstruction-error penalty.
  • Reflectance consistency, measured by RCI at 0.988, becomes a trainable and evaluable target, so low-light methods can be optimized directly for illumination invariance rather than only for pixel fidelity.
  • Replacing RetinexNet's decomposition in CrackNex raises low-light crack segmentation mIoU from 65.6 to 67.6 (and to 68.7 with the wavelet variant) and F1 from 72.3 to 75.4 (and 76.2 with the wavelet variant).
  • Replacing the Retinex module in DAI-Net raises zero-shot day-night detection mAP by 1.2 points on Wider-Face to DarkFace and by 0.8 points on COCO to DarkFace, with the rest of the detector unchanged.
  • On infrared-visible fusion, the quaternion decomposition improves all four reported quality metrics (entropy, spatial frequency, average gradient, standard deviation) over the RDMFuse baseline, indicating better detail and contrast preservation.

Reading between the lines

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

  • Beyond the paper, the RCI could serve as a generic diagnostic for any Retinex-style method, because penalizing the worst pixel across an illumination sweep may catch localized instabilities that average metrics miss.
  • Beyond the paper, the reconstruction loss alone admits trivial factorizations (one factor equal to the input and the other equal to the identity quaternion), so the claimed disentanglement is carried by the reflectance-consistency and mutual-reconstruction losses; ablating those terms while keeping reconstruction would test how much the quaternion algebra itself contributes.
  • Beyond the paper, if the factorization genuinely encodes cross-channel color structure, the same $S=Q_1\otimes Q_2$ scheme could extend to video by enforcing quaternion consistency across frames and to perceptual color spaces aligned with opponent-color processing, directions the paper lists as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript proposes QRetinex-Net, a quaternion-valued Retinex decomposition in which an RGB image S is represented as a purely imaginary quaternion and factorized as S = Q1 ⊗ Q2 under the Hamilton product, with the decomposition performed on Haar wavelet coefficients by a trainable cross-attention network. It introduces the Reflectance Consistency Index (RCI) and reports reconstruction results on LOLv1 (PSNR 62.515 dB, SSIM 0.9998, LPIPS 0.0001, RCI 0.988), as well as gains in low-light crack segmentation, zero-shot day-night object detection, and infrared-visible fusion.

Significance. If the decomposition were mathematically sound and the RCI were independently validated, the quaternion formulation of Retinex would be a useful conceptual contribution, and the breadth of downstream applications is a strength. However, the central factorization is not established: the stated initialization cannot reconstruct the input under the Hamilton product, the loss does not rule out degenerate factorizations, and the RCI is computed with the same decomposition that is explicitly trained to enforce reflectance equality. These issues make the headline numbers uninformative as evidence for a disentangled quaternion Retinex model. The paper does not provide code or machine-checked proofs; as submitted, its value is mainly suggestive.

major comments (4)
  1. [§III-A, Eqs. (4)–(6)] For the stated initialization Q1_init = (0, R/M, G/M, B/M) and Q2_init = (0, R, G, B), the Hamilton product (1) yields scalar part −(R^2+G^2+B^2)/M and zero imaginary part for every pixel. Therefore the RGB image cannot be recovered from the initial quaternion pair by Eq. (3), contradicting the claimed 'perfect reconstruction.' The paper provides no argument that the learned network escapes this zero-imaginary subspace, nor that the reconstruction loss actually enforces equality on the imaginary part of the product.
  2. [§III-C, decomposition loss] Because the input S in Eq. (2) has zero scalar part, any exact factorization S = Q1 ⊗ Q2 forces w1 w2 = v1 · v2. As written, the loss compares the color channels of the product with S, and nothing constrains the scalar channel of Q1 and Q2 individually. The trivial solution Q1 = 1, Q2 = S satisfies exact reconstruction exactly and yields a constant reflectance; L_decomp does not rule it out. This degeneracy also makes RCI vacuous, since a constant reflectance has zero variance and achieves RCI = 1.
  3. [§III-D, RCI and L_equal_R] The RCI is computed using the same decomposition network D that is trained with the reflectance consistency term L_equal_R, which explicitly forces R_low ≈ R_high. Reporting RCI = 0.988 for this network is therefore a check that the training objective was optimized, not an independent index of reflectance stability. In addition, the RCI normalization by the maximum variance 0.25 assumes reflectance values are constrained to [0,1], but no such normalization is imposed on the quaternion reflectance output Q1.
  4. [Table I and Section IV] The reported PSNR of 62.515 dB and SSIM of 0.9998 on LOLv1 are implausibly high for low-light reconstruction and are not directly comparable to the baselines, which optimize different objectives and are not trained to back-reconstruct the input. Furthermore, the abstract's claim of '2–11% performance improvements' is not supported by the tables: Table II shows +2.0 mIoU and +3.1 F1, Table III shows +0.8–1.2 mAP, and Table IV shows only small increments in fusion metrics.
minor comments (4)
  1. [§III-A, notation] The symbol H is used both for the quaternion algebra and for image height; the equations also contain garbled superscripts and subscripts (e.g., 'Q(init∈H%×''), which should be cleaned up.
  2. [Figures 5–6] The caption of Fig. 5 lists 'c) Kind++' while the text refers to subfigure (d) as URetinexNet; the caption and the referenced subfigures are inconsistent, and Fig. 6's caption is truncated.
  3. [Table V] The ablation table reports PSNR/SSIM for 'reconstructed reflectance' and MSE, but the text does not define which image the reconstruction error is measured against; clarify whether this is reflectance, illumination, or the full reconstructed image.
  4. [References] Several references have formatting artifacts (e.g., duplicate 'Lecture Notes in Computer Science' fields, an arXiv citation without a DOI despite a journal version being available), and the claimed 'first Quaternion Retinex formulation' is not positioned against prior quaternion-based color-processing literature beyond a brief mention.

Circularity Check

1 steps flagged · score 6.0 of 10

RCI reflectance-stability claim is optimized into existence by the training loss; downstream benchmarks remain independent.

  1. fitted input called prediction [Section III.C (reflectance consistency loss, item (f)) and Section III.D (RCI metric)]
    "To enforce consistency between reflectances estimated from low- and high-light images, we introduce the reflectance consistency loss: ℒOplYq_r={QRqQs−QRtJet{$ ... An RCI value of 1 corresponds to zero variance and indicates perfectly consistent reflectance across all illumination levels."

    The RCI is computed by applying the trained decomposition D to interpolated low/high-light images and measuring the variance of the reflectance output across illumination levels. But the same network is explicitly trained, via L_equal_R, to minimize the difference between Q_R^low and Q_R^normal. Since the interpolated sequence has endpoints equal to those training images and D is continuous, the variance that RCI penalizes is precisely the objective that the loss already minimizes. Reporting RCI=0.988 as a validation of reflectance stability is therefore reporting the training objective under a new name; the baseline methods in Table I did not optimize this objective.

full rationale

The clearest circular step is the RCI validation. RCI is defined as 1 minus normalized variance of reflectance over illumination interpolation, while the training loss explicitly enforces low-light and high-light reflectance equality. Thus the headline reflectance-consistency number is a re-encoding of the training objective, not an independent prediction. The downstream experiments (crack segmentation, day-night detection, IR-VI fusion) use external benchmarks and unchanged downstream pipelines, so the reported 2-11% gains are not circular. I found no load-bearing self-citation chain: the authors' own prior citation [19] is peripheral, and no uniqueness theorem is invoked to force the quaternion choice. The purely-imaginary initialization and the unconstrained scalar channel of the Hamilton product are serious correctness risks, and the trivial factor Q1=1, Q2=S is not ruled out, but those are failures of derivation rather than reductions of a prediction to its input. Overall, one central metric is constructed by the loss, while the external task improvements remain independent, yielding partial circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central method is a deep network with thousands of learned parameters plus several hand-chosen loss weights. The factorization S = R ⊗ I is a modeling assumption, not a derived property. The RCI metric rests on the assumption that linear interpolation simulates real illumination changes, which is not argued. The rod-cone quaternion response is decorative and unused.

free parameters (5)
  • Decomposition network D weights = learned on LOLv1 (number not disclosed)
    All convolutional, attention, and pointwise filter weights are fitted to minimize the multi-term loss on 256x256 patches; the paper does not state a parameter count or train/test split.
  • Loss weight for mutual reconstruction (0.01) = 0.01
    Hand-chosen coefficient for the mutual reconstruction terms in the total loss.
  • Loss weight for illumination smoothness (0.05) = 0.05
    Hand-chosen coefficient for the smoothness term.
  • Loss weight for reflectance consistency (0.01) = 0.01
    Hand-chosen coefficient for the equal reflectance term that directly enforces low and high light reflectance to match.
  • Loss weight for frequency regularization (0.01) = 0.01
    Hand-chosen coefficient gamma in the frequency-domain regularization loss.
assumptions (4)
  • domain assumption Retinex assumption that an observed pixel S equals a product of reflectance R and illumination I
    Sections II.A and III.A rely on this decomposition; it is ill-posed and not derived.
  • ad hoc to paper Hamilton product of two quaternions can represent a meaningful reflectance/illumination factorization for RGB images
    Section III.A introduces S = R ⊗ I but provides no proof that such a factorization disentangles intrinsic reflectance; the network is trained to approximate it, and the trivial solution Q1 = 1, Q2 = S also satisfies the product.
  • standard math Haar wavelet transform is invertible and preserves all information
    Section III.A uses DWT/IDWT for multiscale decomposition; this is a standard property.
  • domain assumption Linear interpolation between low-light and normal-light images represents a valid illumination change for RCI evaluation
    Section III.D defines RCI over linearly interpolated images; this assumes intermediate images correspond to physically plausible illumination variations, which is not justified.
invented entities (1)
  • Rod-cone quaternion response V = α(Hr) + β(Hc)
    purpose: Motivational framing for quaternion color representation in the introduction and background
    This representation is introduced in Section II.A but never used in the actual decomposition or experiments; it has no independent empirical support and is not connected to the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QRetinex-Net: Quaternion-Valued Retinex Decomposition for Low-Level Computer Vision Applications." pith.science (2026). https://pith.science/paper/CWCXGGXO

@misc{pith2026250716683,
  author       = {Pith},
  title        = {Pith review of: QRetinex-Net: Quaternion-Valued Retinex Decomposition for Low-Level Computer Vision Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWCXGGXO}},
  note         = {Machine review of arXiv:2507.16683}
}
read the original abstract

Images taken in low light often show color shift, low contrast, noise, and other artifacts that hurt computer-vision accuracy. Retinex theory addresses this by viewing an image S as the pixel-wise product of reflectance R and illumination I, mirroring the way people perceive stable object colors under changing light. The decomposition is ill-posed, and classic Retinex models have four key flaws: (i) they treat the red, green, and blue channels independently; (ii) they lack a neuroscientific model of color vision; (iii) they cannot perfectly rebuild the input image; and (iv) they do not explain human color constancy. We introduce the first Quaternion Retinex formulation, in which the scene is written as the Hamilton product of quaternion-valued reflectance and illumination. To gauge how well reflectance stays invariant, we propose the Reflectance Consistency Index. Tests on low-light crack inspection, face detection under varied lighting, and infrared-visible fusion show gains of 2-11 percent over leading methods, with better color fidelity, lower noise, and higher reflectance stability.

Figures

Figures reproduced from arXiv: 2507.16683 by the authors.

Figure 1
Figure 1. Overview of the color vision pipeline and Retinex decomposition. An illumination source with a specific spectral distribution (left) illuminates an object whose surface reflectance selectively reflects certain wavelengths. The reflected light enters the human eye, where rods and three types of cones (L, M, and S) detect overlapping wavelength bands. Retinex processing then decomposes the visual signal into illuminat… view at source ↗
Figure 2
Figure 2. Representative low‐level vision applications leveraging the proposed quaternion‐valued Retinex decomposition framework. Building on this concept, KinD divided the workflow into decomposition, reflectance restoration, and illumination adjustment modules, aided by multi-scale illumination attention [42]. Although KinD improved global and local consistency, it tends to produce overexposed images under severe underexpos… view at source ↗
Figure 3
Figure 3. Quaternion-valued Retinex decomposition framework designed to (i) preserve inter-channel color relationships, (ii) ensure perfect reconstruction, and (iii) learn stable reflectance features across varying illuminations. The core idea is to represent color images as pairs of quaternions in the wavelet domain and to design a network that learns them so that their Hamilton product accurately reconstructs the original i… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: RCI Metric Computation. We linearly interpolate between the low-light and normal-light images to generate intermediate images {Sy}. Each Sz is decomposed by 𝐷 to obtain a reflectance map 𝑅z. The pixel-wise variance of {Rz} is computed across 𝛼, and its supremum over al…
Figure 5
Figure 5. Figure 5: Retinex decomposition (illumination-reflectance) by deep-learning based approaches. Notice the halos and color shifts in the reflectance maps. a) Input image b) RetinexNet c) Kind++ d) U-Retinexnet e) Diff-retinex, f) Quaternion Retinex. Our decomposition results for t…
Figure 6
Figure 6. Figure 6: Variance of the reflectance map with change of the illumination for various methods. The reflectance component retains structural details, and the illumination remains smooth without introducing artifacts. Notably, our reflectance remains stable even under simulated il…
Figure 7
Figure 7. Figure 7: Low-light crack segmentation. a) Input b) CrackNex c) Our method d) Ground truth The primary difference is that the scalar reflectance constraint replaces our quaternion-based constraint. We benchmark the modified CrackNex on the LCSD dataset. As summarized in Table II…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 17 canonical work pages

  1. [1]

    GDIP: Gated differentiable image processing for object detection in adverse conditions,

    S. Kalwar et al., “GDIP: Gated differentiable image processing for object detection in adverse conditions,” in 2023 IEEE International Conference on Robotics and Automation (ICRA), IEEE, May 2023, pp. 7083–7089

    work page 2023

  2. [5]

    Multi-Object Tracking in the Dark,

    X. Wang et al., “Multi-Object Tracking in the Dark,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, openaccess.thecvf.com, 2024, pp. 382–392

    work page 2024

  3. [16]

    Available: http://arxiv.org/abs/2405.03349

    [Online]. Available: http://arxiv.org/abs/2405.03349

    arXiv

  4. [17]

    RetinexFlow for CT metal artifact reduction

    [Online]. Available: http://arxiv.org/abs/2306.10520

    work page Pith review arXiv

  5. [18]

    Color image segmentation in a quaternion framework,

    O. N. Subakan and B. C. Vemuri, “Color image segmentation in a quaternion framework,” Energy Minimization Methods Comput. Vis. Pattern Recognit., vol. 5681, no. 2009, pp. 401–414, Jan

    work page 2009

  6. [20]

    Quaternion wavelet analysis and application in image denoising,

    M. Yin et al., “Quaternion wavelet analysis and application in image denoising,” Math. Probl. Eng., vol. 2012, no. 1, pp. 1–21, Jan

    work page 2012

  7. [21]

    Image processing using the quaternion wavelet transform,

    E. Bayro-Corrochano and M. A. de La Torre Gomora, “Image processing using the quaternion wavelet transform,” in Lecture Notes in Computer Science, in Lecture notes in computer science. , Berlin, Heidelberg: Springer Berlin Heidelberg, 2004, pp. 613–620

    work page 2004

  8. [22]

    Available: http://arxiv.org/abs/1912.01703

    [Online]. Available: http://arxiv.org/abs/1912.01703

    arXiv 1912

Show all 23 references
  1. [23]

    Seeing black and white,

    A. Gilchrist, “Seeing black and white,” 2006, [Online]. Available: https://books.google.com/books?hl=en&lr=&id=0pVr7dhmdWYC&oi=fnd&pg=PR7&dq=A.+Gilchrist,+%22Seeing+black+and+white&ots=kINQOVcdUJ&sig=3ukFfNLUUxKkdTVmn445TecE6EM

    2006

  2. [24]

    Available: http://arxiv.org/abs/2403.03063

    [Online]. Available: http://arxiv.org/abs/2403.03063

    arXiv

  3. [29]

    A weighted variational model for simultaneous reflectance and illumination estimation,

    X. Fu et al., “A weighted variational model for simultaneous reflectance and illumination estimation,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, Jun. 2016, pp. 2782–2790

    2016

  4. [35]

    URetinex-net: Retinex-based deep unfolding network for low-light image enhancement,

    W. Wu et al., “URetinex-net: Retinex-based deep unfolding network for low-light image enhancement,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, Jun

    2022

  5. [38]

    Survey of methods and evaluation of Retinex-inspired image enhancers,

    G. Simone et al., “Survey of methods and evaluation of Retinex-inspired image enhancers,” J. Electron. Imaging, vol. 31, no. 06, Dec. 2022, doi: 10.1117/1.jei.31.6.063055

    2022 doi

  6. [44]

    Zero-shot restoration of underexposed images via robust retinex decomposition,

    A. Zhu et al., “Zero-shot restoration of underexposed images via robust retinex decomposition,” in 2020 IEEE International Conference on Multimedia and Expo (ICME), IEEE, Jul. 2020, pp. 1–6

    2020

  7. [48]

    X-ray image enhancement based on improved Retinex-Net,

    K. Fu et al., “X-ray image enhancement based on improved Retinex-Net,” in 2022 7th International Conference on Automation, Control and Robotics Engineering (CACRE), IEEE, Jul. 2022, pp. 194–198

    2022

  8. [52]

    Image dehazing based on multi-scale retinex and guided filtering,

    Z. Gao and Y. Zhai, “Image dehazing based on multi-scale retinex and guided filtering,” in 2022 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML), IEEE, Oct. 2022, pp. 123–126

    2022

  9. [56]

    Study of enhanced multi-spectral remote-sensing-satellite image technology based on improved retinex-net,

    P. Liu et al., “Study of enhanced multi-spectral remote-sensing-satellite image technology based on improved retinex-net,” in 2022 2nd International Conference on Algorithms, High Performance Computing and Artificial Intelligence (AHPCAI), IEEE, Oct. 2022, pp. 484–489

    2022

  10. [2017]

    Available: http://arxiv.org/abs/1711.05101

    [Online]. Available: http://arxiv.org/abs/1711.05101

    arXiv

  11. [2018]

    Available: http://arxiv.org/abs/1808.04560

    [Online]. Available: http://arxiv.org/abs/1808.04560

    arXiv

  12. [2019]

    doi: 10.1145/3343031.3350926

  13. [2022]

    doi: 10.1109/cvpr52688.2022.00581

    2022

  14. [2023]

    Available: http://arxiv.org/abs/2303.06705

    [Online]. Available: http://arxiv.org/abs/2303.06705

    arXiv

  15. [2024]

    doi: 10.2139/ssrn.4806922

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.