Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Wavelet-based Decoupling Framework for low-light Stereo Image Enhancement

T0 review · 4 major / 5 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 A three-level wavelet decomposition separates illumination from texture, so cross-view interaction can be confined to high-frequency branches and still achieve the best reported enhancement results.

desk verdict A solid, well-ablated engineering paper on wavelet-based decoupling for low-light stereo enhancement; the architecture is new and the ablations are honest, but the dataset-split documentation and the 'first' claim need fixing before the SOTA headline can stand. read the letter →

arxiv 2507.12188 v1 pith:5GG5ZBAK submitted 2025-07-16 cs.CV cs.AIeess.IV

Shuangli Du , Siming Yan , Zhenghao Shi , Zhenzhen You , Lu Sun This is my paper · ORCID
classification cs.CVcs.AIeess.IV
keywords low-lightstereoimageenhancementwavelettransformfeaturedecouplingcross-viewinteractionparallaxattentionilluminationadjustmenttexturedetailrestoration
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 argues that low-light enhancement models fail because they entangle illumination, noise, and texture inside one latent space, and it offers decoupling as the fix. It proposes WDCI-Net, a stereo enhancement network that uses a three-level discrete wavelet transform to split each view into a low-frequency branch for illumination and color correction and three high-frequency branches for texture recovery and denoising. The distinctive design decision is that cross-view interaction happens only on high-frequency features, on the grounds that the other stereo view supplies missing detail but not illumination. The authors train on synthetic stereo pairs with both uniform and non-uniform lighting and report the best PSNR and SSIM on Flickr1024 and KITTI2015 and the best NIQE on real-world Holopix50k among the compared methods. If the paper is right, wavelet-based feature decoupling is a viable alternative to Retinex or Fourier decoupling for stereo low-light enhancement.

What carries the argument

The load-bearing object is the three-level discrete wavelet transform, which decomposes each view's feature map into one low-frequency subband $F_{Li}$ and three directional high-frequency subbands $V_{Li}, H_{Li}, D_{Li}$ at scales $H/2$, $H/4$, and $H/8$, giving lossless downsampling and separate processing paths. The low-frequency branch feeds the Illumination Adjustment Module, which uses large-kernel convolutions, channel attention, and SimpleGate units to correct brightness and color; the high-frequency branches feed HF-CIM, which first fuses the three directions with the Selective Kernel Feature Fusion (SKFF) module, then estimates parallax attention maps between the two views and uses them to transfer the other view's high-frequency features, and finally DTEM, which uses cross-attention guided by the fused feature map to enhance vertical, horizontal, and diagonal details while suppressing noise. The inverse wavelet transform reconstructs the output, and the exchange experiment in Figure 1 is the empirical justification for assigning illumination to the low-frequency branch.

What would settle it

Take a stereo pair with a strong local light source or a severe exposure gradient and run the paper's exchange experiment: swap only the low-frequency component with a normal-light pair and reconstruct. If the reconstructed image retains a visible brightness gradient or color cast carried by the high-frequency bands, or if the PSNR between the swapped reconstruction and the normal-light target drops far below the values reported for LOL pairs, then the claim that illumination is carried by the low-frequency subband fails for non-uniform lighting.

Watch

Extended reading notes

Core claim

The central claim is that illumination lives almost entirely in the low-frequency subband of a multi-level wavelet decomposition, while texture and noise live in the high-frequency subbands, so the two can be enhanced by separate branches and only the high-frequency branches need stereo correspondence. The authors support this with an exchange experiment on 100 pairs from the LOL dataset: swapping the low-frequency component between a low-light image and its normal-light counterpart shifts the reconstructed image's brightness and color toward the source of the low-frequency component. Built on that premise, WDCI-Net's Illumination Adjustment Module (IAM) restores the low-frequency branch under frequency, SSIM, and VGG supervision, while the High-Frequency Guided Cross-view Interaction Module (HF-CIM) uses parallax attention on fused high-frequency features of the two views, and the Detail and Texture Enhancement Module (DTEM) refines those features with cross-attention before inverse wavelet reconstruction. The reported numbers are best among the compared methods on synthetic Flickr1024 and KITTI2015 sets and on the real Holopix50k set under the no-reference NIQE metric.

Load-bearing premise

The whole method rests on the assumption that a three-level wavelet decomposition cleanly separates illumination from texture, so brightness and color can be corrected using only the low-frequency part while high-frequency parts independently handle detail and noise; if extreme or uneven lighting breaks that separation, the architecture's foundation is gone.

Editorial extensions

If this is right

  • Cross-view interaction can be confined to high-frequency branches, so stereo correspondence supplies texture detail while illumination correction stays single-view, reducing the risk of transferring noise between views.
  • Supervising the low-frequency branch as its own sub-task with frequency, SSIM, and VGG losses lets the network adjust illumination and color without compromising detail recovery.
  • Training on both uniform and non-uniform illumination synthetic pairs is what the authors credit for the model's generalization to real-world images, measured by the best average NIQE on Holopix50k.
  • The reported PSNR and SSIM gains over single-image and stereo baselines on Flickr1024 and KITTI2015 imply that wavelet decoupling with high-frequency-only correspondence is a competitive architecture for this task.

Reading between the lines

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

  • The high-frequency-only interaction principle could transfer to other stereo restoration tasks such as super-resolution or deblurring, where the complementary information between views is also concentrated in detail bands rather than in low-frequency content.
  • A natural stress test is to run the exchange experiment on images with strong local light sources or exposure gradients; if the low-frequency subband stops being the sole carrier of illumination, the hard decoupling would need to become an adaptive fusion of the two branches.
  • Because the authors concede that the method still struggles under extreme low-light conditions, replacing the fixed three-level wavelet with a learned or adaptive wavelet is a testable extension that could tune the illumination-texture split per image.
  • The training data are synthetic pairs generated by the LFENet pipeline, so the claimed generalization could be probed further by evaluating on a real paired low-light stereo benchmark with ground truth if one becomes available.
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 / 5 minor

Summary. The paper introduces WDCI-Net, a low-light stereo image enhancement network that decouples illumination and detail restoration by applying a three-level discrete wavelet transform to the feature space. The low-frequency branch is processed by an Illumination Adjustment Module (IAM) with a dedicated sub-supervision loss; high-frequency subbands from both views are fused by a high-frequency cross-view interaction module (HF-CIM) and refined by a Detail and Texture Enhancement Module (DTEM). The model is trained on synthetic uniform and non-uniform low-light stereo pairs generated with the LFENet method from Flickr1024, Holopix50k, and KITTI, and evaluated on synthetic Flickr1024/KITTI test pairs and real Holopix50k pairs. The authors report state-of-the-art PSNR/SSIM and NIQE results and provide code and dataset.

Significance. If the quantitative claims are reproducible, the paper makes a useful contribution: a wavelet-based feature decoupling architecture for stereo low-light enhancement, with cross-view interaction restricted to high-frequency branches, a design choice that is clearly described and ablated. The public code/data release and the module-level ablations (Table III) are strengths. However, the SOTA claim currently rests on comparisons whose validity depends on dataset construction details that are not reported, and the reported margins over DCI-Net are small relative to the absence of variance estimates. With tightened experimental reporting, this could become a solid empirical contribution.

major comments (4)
  1. [Section IV-A.1] The paper does not report a scene-level disjoint split between training and test sets. Training and test are drawn from the same three collections (Flickr1024, Holopix50k, and KITTI), and the test set includes 191 real Holopix50k pairs while the training set contains synthetic Holopix50k pairs from the same collection; Flickr1024 and KITTI similarly appear in both synthetic training and synthetic test sets. Without an explicit statement that scenes are disjoint, the margins in Tables I and II (e.g., PSNR 26.790 vs 26.507 on Flickr left; NIQE 3.2623 vs 3.4014 on Holopix50k) may reflect memorized scene content rather than generalization. The authors should report the number of distinct scenes used for training and testing, state the split criterion, and verify that no scene overlap exists.
  2. [Section IV-A.1] The stated training set size of 1,289 pairs is not consistent with the per-subset counts: 57+459+124+181+296+100=1,217. This discrepancy is not explained in the text. Since the composition of the training set is central to the reported comparisons, the authors should correct the count and provide the exact breakdown, including how the uniform and non-uniform subsets were defined.
  3. [Tables I-III] The quantitative evaluation reports single point estimates of PSNR, SSIM, and NIQE with no error bars, multiple-seed runs, or significance tests. The advantage over the closest competitor DCI-Net is small (roughly 0.28-0.43 dB in PSNR on the synthetic tests and 0.14 in NIQE on Holopix50k), while the ablation differences in Table III are much larger (1.5 dB or more). Without variance information, it is impossible to tell whether the reported gains over DCI-Net are meaningful; the authors should report standard deviations over at least three training runs or provide paired statistical tests on the test set.
  4. [Section I, Figure 1] The central premise that multi-level wavelet decomposition isolates illumination in the low-frequency subband is supported only by the exchange experiment on 100 LOL pairs (Figure 1), and the conclusion concedes limitations under extreme low-light conditions. Given that the training set explicitly includes non-uniform illumination cases, this premise should be demonstrated on non-uniform and real-world examples as well (e.g., low-frequency reconstruction visualizations on Holopix50k); otherwise the decoupling strategy's foundation is not established for the conditions the paper claims to handle.
minor comments (5)
  1. [Section II-B, Section III-E, Table III caption] There are several typographical errors: 'Baed on PAM', 'ground turth', and 'the the results' in the Table III caption.
  2. [Section IV] The dataset is called 'Flickr1024' in Section IV-A.1 but 'Flickr2014' in Section IV-B and the caption of Figure 6; please harmonize the name.
  3. [References] References [11] and [14] are the same paper (Wei et al., Deep Retinex Decomposition), and references [32] and [38] appear to be duplicate entries for the same TMM paper; please deduplicate.
  4. [Equation (9)] The L2 norm is applied to the sum of the left- and right-view VGG differences before the norm, which couples the two views in a way that is not motivated; consider writing the two terms separately or explaining the choice.
  5. [Section IV-D] The ablation discussion repeatedly refers to 'Table II' when the ablation results are in Table III; please correct the cross-references.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the wavelet decoupling design is an empirically motivated architecture choice, and the SOTA claim is tested against external benchmarks rather than derived from fitted parameters.

full rationale

This is an empirical engineering paper with no first-principles derivation chain whose conclusion could collapse into its own assumptions. The central premise—that multi-level wavelet low-frequency components carry illumination and can therefore be adjusted independently of high-frequency detail—is supported by the exchange experiment on 100 LOL pairs (Sec. I, Fig. 1) and then implemented as an architectural inductive bias: IAM processes the low-frequency branch while HF-CIM and DTEM process high-frequency branches. That is a design hypothesis, not a circular prediction. The network is trained with the losses in Eqs. (4)–(9) and evaluated on Flickr1024/KITTI synthetic pairs and real Holopix50k against seven baselines trained on the same data (Sec. IV-B, Tables I–II). The ablations in Table III are empirical sensitivity checks rather than equivalences. Citations to the authors' prior work ([1], [5], [6]) are background or data-generation references; [1] supplies the LFENet procedure for creating synthetic training pairs, which is an input to training rather than a conclusion the paper claims to derive, so it does not make the SOTA result circular. The limitation stated in Sec. V (weak performance under extreme low-light conditions), the numeric mismatch in the training-set counts (1,289 stated vs. 1,217 summed in Sec. IV-A.1), and the absence of a reported scene-level train/test disjointness are correctness and generalization concerns, but none of them is a circularity step: no result reduces by construction, by fitted-parameter renaming, or by a self-citation chain to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The network introduces no new physical entities. The ledger lists the manually chosen hyperparameters and the main domain assumptions that the architecture depends on; all are either standard or not independently validated.

free parameters (3)
  • VGG perceptual loss weight = 0.0001
    Applied in Eq. 9; chosen by hand without reported sensitivity analysis.
  • Initial learning rate = 0.0002
    Set in training; halved every 250 epochs. No ablation or explanation for this value.
  • Number of wavelet decomposition levels = 3
    The architecture uses three-level DWT; no ablation for alternative depths.
assumptions (4)
  • domain assumption Three-level DWT separates illumination into the low-frequency subband and texture/noise into high-frequency subbands
    Motivated by the exchange experiment in Fig. 1, but assumed universally; the authors admit extreme low-light cases are not handled well.
  • ad hoc to paper Cross-view interaction on high-frequency features alone is sufficient for stereo detail recovery
    Design choice; supported only by the HF-CIM ablation, not by analysis or comparison with full-space interaction.
  • domain assumption Synthetic low-light data generated by the LFENet method is representative of real low-light conditions
    Training and synthetic tests rely on this generation; real-world generalization is evaluated via NIQE only.
  • domain assumption Pre-trained VGG19 features are an appropriate perceptual prior for low-frequency illumination maps
    Used in the low-frequency branch loss L_vgg without justification or ablation of the specific weighting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wavelet-based Decoupling Framework for low-light Stereo Image Enhancement." pith.science (2026). https://pith.science/paper/5GG5ZBAK

@misc{pith2026250712188,
  author       = {Pith},
  title        = {Pith review of: Wavelet-based Decoupling Framework for low-light Stereo Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5GG5ZBAK}},
  note         = {Machine review of arXiv:2507.12188}
}
read the original abstract

Low-light images suffer from complex degradation, and existing enhancement methods often encode all degradation factors within a single latent space. This leads to highly entangled features and strong black-box characteristics, making the model prone to shortcut learning. To mitigate the above issues, this paper proposes a wavelet-based low-light stereo image enhancement method with feature space decoupling. Our insight comes from the following findings: (1) Wavelet transform enables the independent processing of low-frequency and high-frequency information. (2) Illumination adjustment can be achieved by adjusting the low-frequency component of a low-light image, extracted through multi-level wavelet decomposition. Thus, by using wavelet transform the feature space is decomposed into a low-frequency branch for illumination adjustment and multiple high-frequency branches for texture enhancement. Additionally, stereo low-light image enhancement can extract useful cues from another view to improve enhancement. To this end, we propose a novel high-frequency guided cross-view interaction module (HF-CIM) that operates within high-frequency branches rather than across the entire feature space, effectively extracting valuable image details from the other view. Furthermore, to enhance the high-frequency information, a detail and texture enhancement module (DTEM) is proposed based on cross-attention mechanism. The model is trained on a dataset consisting of images with uniform illumination and images with non-uniform illumination. Experimental results on both real and synthetic images indicate that our algorithm offers significant advantages in light adjustment while effectively recovering high-frequency information. The code and dataset are publicly available at: https://github.com/Cherisherr/WDCI-Net.git.

Figures

Figures reproduced from arXiv: 2507.12188 by the authors.

Figure 1
Figure 1. (a) Examples of image reconstruction via exchange the low-frequency components; (b) Illustration of PSNR ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of WDCI-Net includes two weight-shared branches for processing the left and right views, respectively. The method comprises [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The detailed process of the the Illumination Adjustment Module [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: (a) The first stage of the HF-SMM module: computing and generating the attention map; (b) The second stage: using the attention map to produce [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The detailed process of the the Detail and Texture Enhancement Module (DTEM). [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the enhanced image of each method on Flickr2014 dataset. It can be seen that compared with other methods, our method effectively [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: The MSE map between ground truth and the enhanced result on the Kitti2015 dataset, where the white area represents large error and the black area [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Visualization of enhanced images on the unsupervised real-world dataset Holopix50. Compared with other methods, our approach achieves superior [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 32 canonical work pages

  1. [1]

    Low-light stereo image enhancement and de-noising in the low-frequency information enhanced image space,

    M. Zhao, X. Qin, S. Du, X. Bai, J. Lyu, and Y . Liu, “Low-light stereo image enhancement and de-noising in the low-frequency information enhanced image space,” Expert Systems with Applications , vol. 265, p. 125803, 2025

    work page 2025

  2. [2]

    Decoupled cross-scale cross-view interaction for stereo image enhancement in the dark,

    H. Zheng, Z. Zhang, J. Fan, R. Hong, Y . Yang, and S. Yan, “Decoupled cross-scale cross-view interaction for stereo image enhancement in the dark,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 1475–1484

    work page 2023

  3. [3]

    A weighted vari- ational model for simultaneous reflectance and illumination estimation,

    X. Fu, D. Zeng, Y . Huang, X.-P. Zhang, and X. Ding, “A weighted vari- ational model for simultaneous reflectance and illumination estimation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2782–2790

    work page 2016

  4. [4]

    Star: A structure and texture aware retinex model,

    J. Xu, Y . Hou, D. Ren, L. Liu, F. Zhu, M. Yu, H. Wang, and L. Shao, “Star: A structure and texture aware retinex model,” IEEE Transactions on Image Processing , vol. 29, pp. 5022–5037, 2020

    work page 2020

  5. [5]

    Low-light image enhancement and denoising via dual-constrained retinex model,

    S. Du, M. Zhao, Y . Liu, Z. You, Z. Shi, J. Li, and Z. Xu, “Low-light image enhancement and denoising via dual-constrained retinex model,” Applied Mathematical Modelling , vol. 116, pp. 1–15, 2023

    work page 2023

  6. [6]

    A new image de- composition approach using pixel-wise analysis sparsity model,

    S. Du, Y . Liu, M. Zhao, Z. Xu, J. Li, and Z. You, “A new image de- composition approach using pixel-wise analysis sparsity model,” Pattern Recognition, vol. 136, p. 109241, 2023

    work page 2023

  7. [7]

    Low-light image enhancement with semi-decoupled decomposition,

    S. Hao, X. Han, Y . Guo, X. Xu, and M. Wang, “Low-light image enhancement with semi-decoupled decomposition,” IEEE transactions on multimedia, vol. 22, no. 12, pp. 3025–3038, 2020

    work page 2020

  8. [8]

    Fourier priors-guided diffusion for zero-shot joint low-light enhance- ment and deblurring,

    X. Lv, S. Zhang, C. Wang, Y . Zheng, B. Zhong, C. Li, and L. Nie, “Fourier priors-guided diffusion for zero-shot joint low-light enhance- ment and deblurring,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 25 378–25 388

    work page 2024

Show all 37 references
  1. [9]

    Fourllie: Boosting low-light image enhancement by fourier frequency information,

    C. Wang, H. Wu, and Z. Jin, “Fourllie: Boosting low-light image enhancement by fourier frequency information,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 7459– 7469

    2023

  2. [10]

    Em- bedding fourier for ultra-high-definition low-light image enhancement. arxiv 2023,

    C. Li, C. Guo, M. Zhou, Z. Liang, S. Zhou, R. Feng, and C. Loy, “Em- bedding fourier for ultra-high-definition low-light image enhancement. arxiv 2023,” arXiv preprint arXiv:2302.11831

    2023 arXiv

  3. [12]

    Fourier space losses for efficient perceptual image super-resolution,

    D. Fuoli, L. Van Gool, and R. Timofte, “Fourier space losses for efficient perceptual image super-resolution,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 2360–2369

    2021

  4. [13]

    Wave-mamba: Wavelet state space model for ultra-high-definition low-light image enhancement,

    W. Zou, H. Gao, W. Yang, and T. Liu, “Wave-mamba: Wavelet state space model for ultra-high-definition low-light image enhancement,” in Proceedings of the 32nd ACM International Conference on Multimedia , 2024, pp. 1534–1543

    2024

  5. [14]

    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

    2018 arXiv

  6. [15]

    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

    2017

  7. [16]

    Low-light image enhancement via a deep hybrid network,

    W. Ren, S. Liu, L. Ma, Q. Xu, X. Xu, X. Cao, J. Du, and M.-H. Yang, “Low-light image enhancement via a deep hybrid network,” IEEE Transactions on Image Processing, vol. 28, no. 9, pp. 4364–4375, 2019

    2019

  8. [17]

    Dslr: Deep stacked laplacian restorer for low- light image enhancement,

    S. Lim and W. Kim, “Dslr: Deep stacked laplacian restorer for low- light image enhancement,” IEEE Transactions on Multimedia , vol. 23, pp. 4272–4284, 2020

    2020

  9. [18]

    Low-light image enhancement via progressive-recursive network,

    J. Li, X. Feng, and Z. Hua, “Low-light image enhancement via progressive-recursive network,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 11, pp. 4227–4240, 2021

    2021

  10. [19]

    Enlightengan: Deep light enhancement without paired supervision,

    Y . Jiang, X. Gong, D. Liu, Y . Cheng, C. Fang, X. Shen, J. Yang, P. Zhou, and Z. Wang, “Enlightengan: Deep light enhancement without paired supervision,” IEEE transactions on image processing, vol. 30, pp. 2340– 2349, 2021

    2021

  11. [20]

    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

    2020

  12. [21]

    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

    2019

  13. [22]

    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

    2022

  14. [23]

    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

    2023

  15. [24]

    Lightendiffusion: Unsupervised low-light image enhancement with latent-retinex diffusion models,

    H. Jiang, A. Luo, X. Liu, S. Han, and S. Liu, “Lightendiffusion: Unsupervised low-light image enhancement with latent-retinex diffusion models,” in European Conference on Computer Vision. Springer, 2024, pp. 161–179

    2024

  16. [25]

    Learning to adapt to light,

    K.-F. Yang, C. Cheng, S.-X. Zhao, H.-M. Yan, X.-S. Zhang, and Y .-J. Li, “Learning to adapt to light,” International Journal of Computer Vision , vol. 131, no. 4, pp. 1022–1041, 2023. 12

    2023

  17. [26]

    Wavelet-based mamba with fourier adjustment for low-light image enhancement,

    J. Tan, S. Pei, W. Qin, B. Fu, X. Li, and L. Huang, “Wavelet-based mamba with fourier adjustment for low-light image enhancement,” in Proceedings of the Asian Conference on Computer Vision , 2024, pp. 3449–3464

    2024

  18. [27]

    Enhancing the spatial resolution of stereo images using a parallax prior,

    D. S. Jeon, S.-H. Baek, I. Choi, and M. H. Kim, “Enhancing the spatial resolution of stereo images using a parallax prior,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 1721–1730

    2018

  19. [28]

    Parallax attention for unsupervised stereo correspondence learning,

    L. Wang, Y . Guo, Y . Wang, Z. Liang, Z. Lin, J. Yang, and W. An, “Parallax attention for unsupervised stereo correspondence learning,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 4, pp. 2108–2125, 2020

    2020

  20. [29]

    Cross view capture for stereo image super-resolution,

    X. Zhu, K. Guo, H. Fang, L. Chen, S. Ren, and B. Hu, “Cross view capture for stereo image super-resolution,” IEEE Transactions on Multimedia, vol. 24, pp. 3074–3086, 2021

    2021

  21. [30]

    Pft-ssr: Parallax fusion trans- former for stereo image super-resolution,

    H. Guo, J. Li, G. Gao, Z. Li, and T. Zeng, “Pft-ssr: Parallax fusion trans- former for stereo image super-resolution,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  22. [31]

    Hybrid transformer and cnn attention network for stereo image super-resolution,

    M. Cheng, H. Ma, Q. Ma, X. Sun, W. Li, Z. Zhang, X. Sheng, S. Zhao, J. Li, and L. Zhang, “Hybrid transformer and cnn attention network for stereo image super-resolution,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 1702–1711

    2023

  23. [32]

    Low-light stereo image enhancement,

    J. Huang, X. Fu, Z. Xiao, F. Zhao, and Z. Xiong, “Low-light stereo image enhancement,” IEEE Transactions on Multimedia , vol. 25, pp. 2978–2992, 2022

    2022

  24. [33]

    Multi-scale interaction network for low-light stereo image enhancement,

    Z. Ji, H. Zheng, Z. Zhang, Q. Ye, Y . Zhao, and M. Xu, “Multi-scale interaction network for low-light stereo image enhancement,” IEEE Transactions on Consumer Electronics , vol. 70, no. 1, pp. 3626–3634, 2023

    2023

  25. [34]

    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,” 2020. [Online]. Available: https://arxiv.org/abs/2003.06792

    2020 arXiv

  26. [35]

    Learning to enhance low-light image via zero-reference deep curve estimation,

    C. Li, C. Guo, and C. C. Loy, “Learning to enhance low-light image via zero-reference deep curve estimation,” 2021. [Online]. Available: https://arxiv.org/abs/2103.00860

    2021 arXiv

  27. [36]

    Ultra-high- definition low-light image enhancement: A benchmark and transformer- based method,

    T. Wang, K. Zhang, T. Shen, W. Luo, B. Stenger, and T. Lu, “Ultra-high- definition low-light image enhancement: A benchmark and transformer- based method,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 3, 2023, pp. 2654–2662

    2023

  28. [37]

    Ecaformer: Low-light image enhancement using cross attention,

    Y . Ruan, H. Ma, W. Li, and X. Wang, “Ecaformer: Low-light image enhancement using cross attention,” 2024. [Online]. Available: https://arxiv.org/abs/2406.13281

    2024 arXiv

  29. [38]

    Low-light stereo image enhancement,

    J. Huang, X. Fu, Z. Xiao, F. Zhao, and Z. Xiong, “Low-light stereo image enhancement,” IEEE Transactions on Multimedia , vol. 25, pp. 2978–2992, 2023

    2023

Pith tools

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