Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

DLEN: Dual Branch of Transformer for Low-Light Image Enhancement in Dual Domains

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read A wavelet-aided, dual-branch transformer network is claimed to set new state-of-the-art scores on low-light image enhancement benchmarks.

desk verdict A plausible incremental Retinexformer variant whose SOTA claim rests on a 0.01 dB edge and is undercut by an ablation table that beats the full model on one dataset. read the letter →

arxiv 2501.12235 v4 pith:PMGKBGQI submitted 2025-01-21 cs.CV eess.IV

classification cs.CVeess.IV
keywords low-lightimageenhancementlearnablewavelettransformRetinextheoryTransformerattentionilluminationestimationdual-branchrestorerLOLdatasetPSNRandSSIM
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

Low-light images lose brightness, contrast, and fine detail, and the paper argues that the fix should happen in both the spatial and frequency domains at once. It proposes DLEN, a network whose illumination-estimation stage contains a learnable wavelet transform that separates high- and low-frequency features, followed by a dual-branch Transformer restorer: one branch uses illumination-guided attention to brighten dark regions, the other repairs structure. On the standard LOL benchmarks, DLEN reports 23.942 dB PSNR / 0.841 SSIM on LOLv1 and 22.230 dB PSNR / 0.854 SSIM on LOLv2-real, the highest figures among the methods compared, with a 0.01 dB lead over the best prior one-stage Retinex transformer on LOLv1 and about a 1 dB lead on LOLv2-real. If the claim stands, it means combining frequency-domain separation with illumination-aware attention is a practical recipe for preserving texture while correcting exposure.

What carries the argument

The load-bearing object is the Learnable Wavelet Module (LWN), a trainable 1D filter pair $\vec{h}_0$ and $\vec{h}_1$ whose outer products form the four 2D subband filters $G_{ll}$, $G_{lh}$, $G_{hl}$, and $G_{hh}$, so a feature map is split into low- and high-frequency components before restoration. Two attention blocks carry the dual-branch restorer: the Multi-Illumination Attention Block (MIAB), where the illumination feature $Y$ scales the values before softmax attention, and the Structure Enhancement Attention Block (SEAB), where pointwise and depthwise convolutions enrich queries, keys, and values and the illumination map gates the output. Together these pieces turn the Retinex decomposition and wavelet subbands into residual corrections that preserve texture and structure.

What would settle it

Take the trained filters $\vec{h}_0$ and $\vec{h}_1$ and compute the four subband outputs on an image with smooth regions; if the high-frequency subbands are not near zero on those regions, or if the pair fails a quadrature-mirror condition such as $h_1[n] = (-1)^n h_0[N-1-n]$, then the module is not performing a wavelet separation and the frequency-domain rationale is unsupported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a low-light enhancement network can be made state-of-the-art by inserting a learnable discrete wavelet module into the illumination estimator and splitting the restorer into two Transformer-based branches. The Light Component Predictor decomposes the input under a perturbed Retinex model $I = (R + \tilde{R}) \odot (L + \tilde{L})$, produces an illuminated image $I^{lu}$ and an illumination feature $F^{lu}$; the learnable wavelet then splits $F^{lu}$ into low- and high-frequency subbands before the restorer. The restorer's illumination branch modulates self-attention with the illumination feature via $(Y \odot V)\,\mathrm{softmax}(K^{T}Q/\alpha)$, while the structure branch applies depthwise-convolution-enriched attention and multiplies by the illumination map. Combined, the branches output residuals added to $I^{lu}$, and the network is trained end-to-end with mean absolute error. As reported in Table I, the result is the top PSNR/SSIM among all compared supervised methods on both LOLv1 and LOLv2-real.

Load-bearing premise

The frequency-domain story rests on the assumption that the two learned 1D filters still behave as a genuine low-pass and high-pass wavelet pair after training, but nothing in the training procedure forces them to.

Editorial extensions

If this is right

  • If the reported scores hold, DLEN becomes the top-scoring compared method on LOLv1 and LOLv2-real, so downstream tasks such as detection and recognition on enhanced images inherit a modest quantitative improvement.
  • The ablation numbers attribute roughly 0.48 dB on LOLv1 to the learnable wavelet module and roughly 1.1 dB to the structure branch, implying each component carries part of the gain.
  • Because the learnable wavelet is a frequency-domain front end, it can be attached to other Retinex-based illumination estimators without redesigning the restorer.
  • Training with only mean absolute error and standard augmentation is enough to reach these scores, suggesting the architecture, not a bespoke loss, is doing the work.

Reading between the lines

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

  • Editorial inference: the learned filters are never constrained to satisfy orthogonality or perfect-reconstruction conditions, so at convergence they may act as ordinary separable convolutions; if so, the LWN's benefit would come from extra capacity rather than true frequency separation.
  • Editorial inference: a direct test would replace the learned filter pair with fixed Haar filters and with free $1 \times 1$ convolution kernels while holding everything else fixed; the paper's reasoning predicts the learned pair should beat both, but this experiment is not run.
  • Editorial inference: the same dual-branch, prior-guided attention design could transfer to other restoration tasks with spatially varying conditions, such as underwater enhancement or shadow removal, where a prior map plays the role of the illumination feature.
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.

Referee Report

4 major / 4 minor

Summary. The manuscript proposes DLEN, a low-light image enhancement network that combines a Light Component Predictor with a learnable wavelet module and a Dual-Branch Restorer containing two attention mechanisms (MIAB and SEAB). The paper claims state-of-the-art performance on the LOLv1 and LOLv2-real benchmarks based on PSNR and SSIM comparisons in Table I, and it presents ablation studies in Table II as well as qualitative results. The central claim, stated in the abstract and introduction, is that DLEN outperforms existing methods through extensive experiments.

Significance. If the reported results are reproducible and the internal inconsistency in the ablation study is resolved, DLEN would be a competitive architecture for low-light image enhancement, although its measured margin over Retinexformer on LOLv1 is only 0.01 dB and it trails MIRNet in PSNR on that dataset. The combination of illumination-guided attention with a frequency-domain module is a plausible and interesting direction. The paper gives a clear architectural description and provides a code link, which are strengths; however, the evidence as presented is not sufficient to establish the claimed superiority, and one ablation row directly contradicts the main text.

major comments (4)
  1. [Table I and Section IV-C] The claim that DLEN 'outperforms state-of-the-art methods on standard benchmarks' is not supported by Table I as reported. On LOLv1, MIRNet achieves a higher PSNR (24.140 dB) than DLEN (23.942 dB), and the margin over Retinexformer is only 0.01 dB PSNR and 0.01 SSIM on a 15-image test set, with no error bars, repeated runs, or statistical significance tests. Section IV-C reports only average gains across the two datasets, which conceals the LOLv1 deficit relative to MIRNet. The paper should report per-dataset and per-baseline results with uncertainty quantification, and the abstract and contribution claims should be qualified accordingly.
  2. [Table II and Section IV-D] The ablation study contradicts the main text: removing the Learnable Wavelet Network improves LOLv2-real PSNR from 22.230 to 22.293 dB, while Section IV-D states that 'our full setup achieves the highest PSNR and SSIM values.' This internal inconsistency must be addressed, either by explaining the variation (for example, training noise) or by re-running the ablation, and it weakens the causal claim that the LWN is beneficial as claimed.
  3. [Section III-A, Eq. (4)] The simplification from Eq. (3) to Eq. (4), written as I_lu = I ⊙ L~ = R + C, is not derived, and it is not evident that the left-hand side follows from Eq. (3) for any definition of L~ given in the text. The manuscript should provide the missing derivation or restate the equation with an explicit definition of C; as written, this step is a gap in the formulation of the network in Eqs. (5)-(7).
  4. [Section III-C, Eqs. (19)-(22)] The learnable filters h0 and h1 are used to define a wavelet filter bank, but the paper imposes no orthogonality, reconstruction, or normalization constraint on them, so after training there is no guarantee that they separate low- and high-frequency components as claimed. This undermines the frequency-domain rationale for the LWN as a 'learnable wavelet transform.' The authors should either add constraints or demonstrate empirically that the learned filters behave as a filter bank, for example by visualizing their frequency responses or the resulting subbands.
minor comments (4)
  1. [Eq. (6) and Section III-B] The symbols ILB and SEB are introduced in Eq. (6), but the text later uses SEAB for the structure enhancement branch; a single consistent naming convention should be used throughout.
  2. [References] Several entries in the reference list, including [38] through [53], do not appear to be cited in the body of the paper, and some citations such as [7] omit co-authors; the bibliography should be reconciled with the in-text citations.
  3. [Section IV-B] The maximum pixel value is denoted R in the PSNR formula, which collides with the reflection image R defined in Eq. (1); a different symbol such as P_max would avoid the ambiguity.
  4. [Section IV-A] The implementation section reports image resolution and batch sizes but not the total number of parameters, FLOPs, or training epochs; including these would allow readers to assess the computational overhead of the dual-branch architecture.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DLEN's benchmark gains are measured on held-out test data, and the only overlapping-author citation (RetinexMamba [37]) is not load-bearing.

full rationale

The central claim that DLEN outperforms state-of-the-art methods is supported by Table I, which reports PSNR/SSIM computed directly on the LOL test sets; these numbers are measurements of the trained network on held-out pairs, not quantities derived from the network's own fitted parameters. No equation in Section III evaluates to the reported metrics by construction, and the ablation study compares trained configurations rather than re-inserting fitted values as predictions. The learnable wavelet module in Section III-C is described via Eqs. (19)-(22), but its claimed benefit is an architectural assertion, not a result that reduces to its own definition. The only self-citation is RetinexMamba [37], co-authored by Jiesong Bai, and it appears solely in the survey sentence 'transformer models, and mamba [37], have set new benchmarks'; it is not invoked to justify the DLEN design, the wavelet module, or the benchmark results. An internal inconsistency does exist in Table II, where 'w/o LWN' beats the full model on LOLv2-real PSNR (22.293 vs 22.230), and the LOLv1 margin over Retinexformer is only 0.01 dB without error bars; these are correctness/reproducibility concerns, not circularity. Therefore no load-bearing step reduces to its own input, and the circularity score is 0.

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

The paper's theoretical foundations are a standard Retinex decomposition plus two ad hoc constructs: an unexplained perturbation simplification (Eq. 4) and unconstrained 'learnable wavelet' filters. The trained network also contains learnable attention scales and all ordinary weights, which are fitted to the training data. There are no externally calibrated constants, so the empirical comparison is not circular, but the method's claimed frequency-domain mechanism is not guaranteed by the architecture.

free parameters (4)
  • alpha_i (MIAB attention scale, Eq. 11)
    A per-head learnable scalar that scales K^T Q before softmax; it is fitted on LOL training data and its final value is not reported.
  • beta_F (SEAB attention scale, Eq. 15)
    A learnable scalar controlling the dot product magnitude in structure attention; fitted during training, value not reported.
  • h0, h1 (learnable wavelet filters, Eqs. 19-22)
    Low-pass and high-pass 1D filters are trained instead of being fixed Haar filters; final coefficients are not reported and no orthogonality constraint is applied.
  • training hyperparameters = 128x128 patches, batch sizes 8/4, Adam, MAE loss
    Chosen by hand; these affect reported metrics but are not ablated.
assumptions (4)
  • domain assumption An image can be decomposed as I = R ⊙ L (reflection and illumination).
    I = R ⊙ L is the starting model (Eq. 1) and is accepted without validation in this paper.
  • ad hoc to paper The expansion in Eq. (3) can be simplified to I_lu = I ⊙ L~ = R + C, with C representing perturbation losses.
    Eq. (3) to (4) asserts this simplification; no derivation is given and C is undefined, yet it motivates the full pipeline.
  • ad hoc to paper Trained filter coefficients h0, h1 still function as an orthogonal low/high-pass filter bank for frequency separation.
    The module assumes trained filters still separate low and high frequencies as in Eqs. (19) to (22); no constraint or proof is supplied.
  • domain assumption The per-channel mean value Lp is a useful illumination prior.
    Introduced in Section III-A; plausible but not evaluated in the ablations.
invented entities (2)
  • Perturbation terms R-tilde and L-tilde
    purpose: To extend the Retinex model I = R ⊙ L so that noise and uneven lighting are represented separately.
    Introduced in Eq. (2); they are not measured, visualized, or given any falsifiable prediction, so they are free modeling devices.
  • Perturbation loss term C
    purpose: To patch the undefined algebra in Eq. (4) by claiming perturbation losses collect into a single term C.
    C is never defined, computed, or used, making it an invented placeholder rather than a testable quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DLEN: Dual Branch of Transformer for Low-Light Image Enhancement in Dual Domains." pith.science (2026). https://pith.science/paper/PMGKBGQI

@misc{pith2026250112235,
  author       = {Pith},
  title        = {Pith review of: DLEN: Dual Branch of Transformer for Low-Light Image Enhancement in Dual Domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PMGKBGQI}},
  note         = {Machine review of arXiv:2501.12235}
}
read the original abstract

Low-light image enhancement (LLE) aims to improve the visual quality of images captured in poorly lit conditions, which often suffer from low brightness, low contrast, noise, and color distortions. These issues hinder the performance of computer vision tasks such as object detection, facial recognition, and autonomous driving.Traditional enhancement techniques, such as multi-scale fusion and histogram equalization, fail to preserve fine details and often struggle with maintaining the natural appearance of enhanced images under complex lighting conditions. Although the Retinex theory provides a foundation for image decomposition, it often amplifies noise, leading to suboptimal image quality. In this paper, we propose the Dual Light Enhance Network (DLEN), a novel architecture that incorporates two distinct attention mechanisms, considering both spatial and frequency domains. Our model introduces a learnable wavelet transform module in the illumination estimation phase, preserving high- and low-frequency components to enhance edge and texture details. Additionally, we design a dual-branch structure that leverages the power of the Transformer architecture to enhance both the illumination and structural components of the image.Through extensive experiments, our model outperforms state-of-the-art methods on standard benchmarks.Code is available here: https://github.com/LaLaLoXX/DLEN

Figures

Figures reproduced from arXiv: 2501.12235 by the authors.

Figure 1
Figure 1. The target image (left) and the enhanced result (right) produced by [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualized results on a benchmark dataset. Each image shows the input image, the output produced by our method, and the corresponding target [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The figure illustrates the detailed structure of our model, which consists of two main components: (a) the Light Component Predictor and (b) the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The figure shows the qualitative experimental results on LOLv1. Our method effectively reduces color distortions and enhances lighting effects. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: This figure demonstrates the detailed effects of the module on the [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: The above shows the qualitative experimental results on LOLv2-real. Our method effectively reduces color distortions and enhances lighting effects. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CMAMRNet: A Contextual Mask-Aware Network Enhancing Mural Restoration Through Comprehensive Mask Guidance

    cs.CV 2025-08 conditional novelty 5.0 of 10

    A mask-aware transformer U-Net with mask-guided up/down-samplers and a co-feature aggregator reports state-of-the-art mural restoration on MuralDH and Dunhuang benchmarks.

Reference graph

Works this paper leans on

55 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [31]

    Convo- lutional and fourier neural networks for speckle denoising of wrapped phase in digital holographic interferometry,

    Q. Fang, Q. Li, Q. Song, S. Montresor, P. Picart, and H. Xia, “Convo- lutional and fourier neural networks for speckle denoising of wrapped phase in digital holographic interferometry,” Optics Communications , vol. 550, p. 129955, 2024

  2. [1]

    A simple and effective histogram equalization approach to image enhancement,

    H.-D. Cheng and X. Shi, “A simple and effective histogram equalization approach to image enhancement,” Digital signal processing , vol. 14, no. 2, pp. 158–170, 2004

  3. [2]

    A dynamic histogram equalization for image contrast enhancement,

    M. Abdullah-Al-Wadud, M. H. Kabir, M. A. A. Dewan, and O. Chae, “A dynamic histogram equalization for image contrast enhancement,” IEEE transactions on consumer electronics , vol. 53, no. 2, pp. 593–600, 2007

  4. [3]

    Contrast enhancement based on layered difference representation of 2d histograms,

    C. Lee, C. Lee, and C.-S. Kim, “Contrast enhancement based on layered difference representation of 2d histograms,” IEEE transactions on image processing, vol. 22, no. 12, pp. 5372–5384, 2013

  5. [4]

    Lightness and retinex theory,

    E. H. Land and J. J. McCann, “Lightness and retinex theory,” Josa, vol. 61, no. 1, pp. 1–11, 1971

  6. [5]

    Mbllen: Low-light image/video enhancement using cnns

    F. Lv, F. Lu, J. Wu, and C. Lim, “Mbllen: Low-light image/video enhancement using cnns.” in BMVC, vol. 220, no. 1. Northumbria University, 2018, p. 4

  7. [6]

    Retinex- former: One-stage retinex-based transformer for low-light image en- hancement,

    Y . Cai, H. Bian, J. Lin, H. Wang, R. Timofte, and Y . Zhang, “Retinex- former: One-stage retinex-based transformer for low-light image en- hancement,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 12 504–12 513

  8. [7]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

Show all 55 references
  1. [8]

    Degradation-aware unfolding half-shuffle transformer for spectral compressive imaging,

    Y . Cai, J. Lin, H. Wang, X. Yuan, H. Ding, Y . Zhang, R. Timofte, and L. V . Gool, “Degradation-aware unfolding half-shuffle transformer for spectral compressive imaging,” Advances in Neural Information Processing Systems, vol. 35, pp. 37 749–37 761, 2022

  2. [9]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739

  3. [10]

    Beyond brightening low-light images,

    Y . Zhang, X. Guo, J. Ma, W. Liu, and J. Zhang, “Beyond brightening low-light images,” International Journal of Computer Vision , vol. 129, pp. 1013–1037, 2021

  4. [11]

    Sparse gradient reg- ularized deep retinex network for robust low-light image enhancement,

    W. Yang, W. Wang, H. Huang, S. Wang, and J. Liu, “Sparse gradient reg- ularized deep retinex network for robust low-light image enhancement,” IEEE Transactions on Image Processing , vol. 30, pp. 2072–2086, 2021

  5. [12]

    An adaptive gamma correction for image enhance- ment,

    S. Rahman, M. M. Rahman, M. Abdullah-Al-Wadud, G. D. Al-Quaderi, and M. Shoyaib, “An adaptive gamma correction for image enhance- ment,” EURASIP Journal on Image and Video Processing , vol. 2016, pp. 1–13, 2016

  6. [13]

    Retinex processing for automatic image enhancement,

    Z.-u. Rahman, D. J. Jobson, and G. A. Woodell, “Retinex processing for automatic image enhancement,” Journal of Electronic imaging , vol. 13, no. 1, pp. 100–110, 2004

  7. [14]

    Naturalness preserved enhancement algorithm for non-uniform illumination images,

    S. Wang, J. Zheng, H.-M. Hu, and B. Li, “Naturalness preserved enhancement algorithm for non-uniform illumination images,” IEEE transactions on image processing , vol. 22, no. 9, pp. 3538–3548, 2013

  8. [15]

    Deep retinex decomposition for low-light enhancement,

    C. Wei, W. Wang, W. Yang, and J. Liu, “Deep retinex decomposition for low-light enhancement,” arXiv preprint arXiv:1808.04560 , 2018

  9. [16]

    Kindling the darkness: A practical low-light image enhancer,

    Y . Zhang, J. Zhang, and X. Guo, “Kindling the darkness: A practical low-light image enhancer,” inProceedings of the 27th ACM international conference on multimedia , 2019, pp. 1632–1640

  10. [17]

    Lime: Low-light image enhancement via illumination map estimation,

    X. Guo, Y . Li, and H. Ling, “Lime: Low-light image enhancement via illumination map estimation,” IEEE Transactions on image processing , vol. 26, no. 2, pp. 982–993, 2016

  11. [18]

    Llnet: A deep autoencoder approach to natural low-light image enhancement,

    K. G. Lore, A. Akintayo, and S. Sarkar, “Llnet: A deep autoencoder approach to natural low-light image enhancement,” Pattern Recognition, vol. 61, pp. 650–662, 2017

  12. [19]

    Learning enriched features for real image restoration and enhancement,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Learning enriched features for real image restoration and enhancement,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXV

  13. [20]

    Springer, 2020, pp. 492–511

  14. [21]

    Zero- reference deep curve estimation for low-light image enhancement,

    C. Guo, C. Li, J. Guo, C. C. Loy, J. Hou, S. Kwong, and R. Cong, “Zero- reference deep curve estimation for low-light image enhancement,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 1780–1789

  15. [22]

    Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement,

    R. Liu, L. Ma, J. Zhang, X. Fan, and Z. Luo, “Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 10 561–10 570

  16. [23]

    Uretinex- net: Retinex-based deep unfolding network for low-light image enhance- ment,

    W. Wu, J. Weng, P. Zhang, X. Wang, W. Yang, and J. Jiang, “Uretinex- net: Retinex-based deep unfolding network for low-light image enhance- ment,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 5901–5910

  17. [24]

    Snr-aware low-light image enhancement,

    X. Xu, R. Wang, C.-W. Fu, and J. Jia, “Snr-aware low-light image enhancement,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 17 714–17 724

  18. [25]

    Toward fast, flexible, and robust low-light image enhancement,

    L. Ma, T. Ma, R. Liu, X. Fan, and Z. Luo, “Toward fast, flexible, and robust low-light image enhancement,” 2022

  19. [26]

    Implicit neural repre- sentation for cooperative low-light image enhancement,

    S. Yang, M. Ding, Y . Wu, Z. Li, and J. Zhang, “Implicit neural repre- sentation for cooperative low-light image enhancement,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 12 918–12 927

  20. [27]

    Learning a simple low-light image enhancer from paired low-light instances,

    Z. Fu, Y . Yang, X. Tu, Y . Huang, X. Ding, and K.-K. Ma, “Learning a simple low-light image enhancer from paired low-light instances,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 252–22 261

  21. [28]

    Iterative prompt learning for unsupervised backlit image enhancement,

    Z. Liang, C. Li, S. Zhou, R. Feng, and C. C. Loy, “Iterative prompt learning for unsupervised backlit image enhancement,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 8094–8103

  22. [29]

    Exploiting deep generative prior for versatile image restoration and manipulation,

    X. Pan, X. Zhan, B. Dai, D. Lin, C. C. Loy, and P. Luo, “Exploiting deep generative prior for versatile image restoration and manipulation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 11, pp. 7474–7489, 2021

  23. [30]

    Improved wavelet threshold for image de-noising,

    Y . Zhang, W. Ding, Z. Pan, and J. Qin, “Improved wavelet threshold for image de-noising,” Frontiers in neuroscience , vol. 13, p. 39, 2019

  24. [32]

    Efficient multi-scale network with learnable discrete wavelet transform for blind motion deblurring,

    X. Gao, T. Qiu, X. Zhang, H. Bai, K. Liu, X. Huang, H. Wei, G. Zhang, and H. Liu, “Efficient multi-scale network with learnable discrete wavelet transform for blind motion deblurring,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024,...

  25. [33]

    Single image dehazing by multi-scale fusion,

    C. O. Ancuti and C. Ancuti, “Single image dehazing by multi-scale fusion,” IEEE Transactions on Image Processing , vol. 22, no. 8, pp. 3271–3282, 2013

  26. [34]

    Multi-scale progressive fusion network for single image deraining,

    K. Jiang, Z. Wang, P. Yi, C. Chen, B. Huang, Y . Luo, J. Ma, and J. Jiang, “Multi-scale progressive fusion network for single image deraining,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 8346–8355

  27. [35]

    An enhancement method for x-ray image via fuzzy noise removal and homomorphic filtering,

    L. Xiao, C. Li, Z. Wu, and T. Wang, “An enhancement method for x-ray image via fuzzy noise removal and homomorphic filtering,” Neurocomputing, vol. 195, pp. 56–64, 2016

  28. [36]

    Image enhancement by stochastic homo- morphic filtering,

    R. Fries and J. Modestino, “Image enhancement by stochastic homo- morphic filtering,” IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 27, no. 6, pp. 625–637, 1979

  29. [37]

    Image super-resolution using deep convolutional networks,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Image super-resolution using deep convolutional networks,” IEEE transactions on pattern analysis and machine intelligence , vol. 38, no. 2, pp. 295–307, 2015

  30. [38]

    Retinexmamba: Retinex-based mamba for low-light image enhancement,

    J. Bai, Y . Yin, Q. He, Y . Li, and X. Zhang, “Retinexmamba: Retinex-based mamba for low-light image enhancement,” arXiv preprint arXiv:2405.03349, 2024

  31. [39]

    High-fidelity document stain removal via a large-scale real- world dataset and a memory-augmented transformer,

    M. Li, H. Sun, Y . Lei, X. Zhang, Y . Dong, Y . Zhou, Z. Li, and X. Chen, “High-fidelity document stain removal via a large-scale real- world dataset and a memory-augmented transformer,” in Proc. WACV, 2024

  32. [40]

    Underwater image restoration through a prior guided hybrid sense approach and extensive benchmark analysis,

    X. Guo, X. Chen, S. Wang, and C.-M. Pun, “Underwater image restoration through a prior guided hybrid sense approach and extensive benchmark analysis,” IEEE TCSVT , 2025

  33. [41]

    High-resolution document shadow removal via a large-scale real-world dataset and a frequency- aware shadow erasing net,

    Z. Li, X. Chen, C.-M. Pun, and X. Cun, “High-resolution document shadow removal via a large-scale real-world dataset and a frequency- aware shadow erasing net,” in Proc. ICCV, 2023, pp. 12449–12458

  34. [42]

    Dual-hybrid attention network for specular highlight removal,

    X. Guo, X. Chen, S. Luo, S. Wang, and C.-M. Pun, “Dual-hybrid attention network for specular highlight removal,” in Proc. ACM MM , 2024, pp. 10173–10181

  35. [43]

    Underwater image restoration via polymorphic large kernel CNNs,

    X. Guo, Y . Dong, X. Chen, W. Chen, Z. Li, F. Zheng, and C.-M. Pun, “Underwater image restoration via polymorphic large kernel CNNs,” in Proc. ICASSP, 2025, pp. 1–5

  36. [44]

    Docdeshadower: Frequency-aware transformer for document shadow removal,

    Z. Zhou, Y . Lei, X. Chen, S. Luo, W. Zhang, C.-M. Pun, and Z. Wang, “Docdeshadower: Frequency-aware transformer for document shadow removal,” in Proc. SMC, 2024, pp. 2468–2473

  37. [45]

    SMAFormer: Synergistic multi-attention transformer for medical image segmentation,

    F. Zheng, X. Chen, W. Liu, H. Li, Y . Lei, J. He, C.-M. Pun, and S. Zhou, “SMAFormer: Synergistic multi-attention transformer for medical image segmentation,” in Proc. BIBM, 2024, pp. 4048–4053

  38. [46]

    Brain diffuser: An end- to-end brain image to brain network pipeline,

    X. Chen, B. Lei, C.-M. Pun, and S. Wang, “Brain diffuser: An end- to-end brain image to brain network pipeline,” in Proc. PRCV , 2023, pp. 16–26

  39. [47]

    Devi- gnet: High-resolution vignetting removal via a dual aggregated fusion transformer with adaptive channel expansion,

    S. Luo, X. Chen, W. Chen, Z. Li, S. Wang, and C.-M. Pun, “Devi- gnet: High-resolution vignetting removal via a dual aggregated fusion transformer with adaptive channel expansion,” in Proc. AAAI , 2024, pp. 4000–4008

  40. [48]

    A large-scale film style dataset for learning multi-frequency driven film enhancement,

    Z. Li, X. Chen, S. Wang, and C.-M. Pun, “A large-scale film style dataset for learning multi-frequency driven film enhancement,” in Proc. IJCAI, 2023, pp. 1160–1168

  41. [49]

    Medprompt: Cross-modal prompting for multi-task medical image translation,

    X. Chen, C.-M. Pun, and S. Wang, “Medprompt: Cross-modal prompting for multi-task medical image translation,” in Proc. PRCV, 2024, pp. 61– 75

  42. [50]

    Test-time intensity consistency adaptation for shadow detection,

    L. Zhu, W. Liu, X. Chen, Z. Li, X. Chen, Z. Wang, and C.-M. Pun, “Test-time intensity consistency adaptation for shadow detection,” arXiv, 2024

  43. [51]

    IMAN: An adaptive network for robust NPC mortality prediction with missing modalities,

    Y . Huo, G. Huang, L. Cheng, J. He, X. Chen, X. Yuan, G. Zhong, and C.-M. Pun, “IMAN: An adaptive network for robust NPC mortality prediction with missing modalities,” in Proc. BIBM , 2024, pp. 2074– 2079

  44. [52]

    FAQNet: Frequency-aware quaternion network for endoscopic highlight removal,

    D. Zhu, G. Huang, X. Yuan, X. Chen, G. Zhong, C.-M. Pun, and J. Deng, “FAQNet: Frequency-aware quaternion network for endoscopic highlight removal,” in Proc. BIBM, 2024, pp. 1408–1413

  45. [53]

    Sienet: Siamese expansion network for image extrapolation,

    X. Zhang, F. Chen, C. Wang, M. Tao, and G.-P. Jiang, “Sienet: Siamese expansion network for image extrapolation,” IEEE Signal Processing Letters, vol. 27, pp. 1590–1594, 2020

  46. [54]

    SpA-Former: An effective and lightweight transformer for image shadow removal,

    X. Zhang, Y . Zhao, C. Gu, C. Lu, and S. Zhu, “SpA-Former: An effective and lightweight transformer for image shadow removal,” inProc. IJCNN, 2023, pp. 1–8

  47. [55]

    MuralDiff: Diffusion for ancient murals restoration on large-scale pre-training,

    Z. Xu, X. Zhang, W. Chen, J. Liu, T. Xu, and Z. Wang, “MuralDiff: Diffusion for ancient murals restoration on large-scale pre-training,” IEEE Transactions on Emerging Topics in Computational Intelligence , 2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.