Multi-Knowledge-oriented Nighttime Haze Imaging Enhancer for Vision-driven Intelligent Systems

Ai Chen; Dong Yang; Duanbing Chen; Junlin Zhou; Yan Fu; Yuxu Lu

REVIEW 4 major objections 5 minor 52 references

Multi-Knowledge-oriented Nighttime Haze Imaging Enhancer for Vision-driven Intelligent Systems

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

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

Pith's one-line read A single network tops task-specific models for haze and low light.

desk verdict A practical all-in-one enhancer with a clean architecture, but the reported superiority margins rest on an evaluation protocol that the paper never fully discloses. read the letter →

arxiv 2502.07351 v4 pith:LCZZXME3 submitted 2025-02-11 cs.CV cs.AI

Ai Chen , Yuxu Lu , Dong Yang , Junlin Zhou , Yan Fu , Duanbing Chen This is my paper · ORCID

classification cs.CVcs.AI

keywords nighttimehazeenhancementlow-lightimagedehazingsalientobjectdetectionmulti-tasklearningself-attentionmulti-receptive-fieldrestoration

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

The paper proposes a single image-restoration network, MKoIE, that aims to handle three degradations at once: daytime haze, low light, and nighttime haze. Its central claim is that one multi-task enhancer beats methods built for each task individually, in both pixel-level reconstruction quality and downstream salient-object detection. The authors argue this matters for vision-driven intelligent systems, where a single lightweight preprocessing block could replace several restoration modules and support reliable object detection in changing conditions. If the claim holds, MKoIE would be a practical all-in-one front end for real-time imaging pipelines.

What carries the argument

The load-bearing components are the task-oriented node learning mechanism, which assigns each degradation type its own small encoder-decoder subnode over a shared backbone; the self-attention module, which lets the nighttime subnode attend to features from the low-light and daytime-dehazing subnodes; and the multi-receptive-field enhancement module, which runs three parallel depthwise separable convolutions with dilation rates 1, 3, and 5 to capture multi-scale features cheaply. A hybrid loss combining L1, Charbonnier, and VGG-based perceptual terms guides training.

What would settle it

Run MKoIE and all compared baselines on a previously unseen test set drawn from RESIDE and CDD, with baselines retrained on the identical training split and MKoIE evaluated only on images it did not train on; if the PSNR, SSIM, and NIQE margins shrink to within noise or reverse, the superiority claim is not established. A simpler check is verifying that no CDD test image also appears among the 3183 training patches.

Watch

Extended reading notes

Core claim

MKoIE is a multi-task encoder-decoder that shares one backbone and splits into three task-oriented subnodes for daytime dehazing, low-light enhancement, and nighttime haze enhancement. For the nighttime task it fuses the outputs of the other two subnodes through a learnable weight and self-attention, treating low-light and daytime-dehazing knowledge as auxiliary cues. The paper reports that MKoIE achieves the highest average PSNR and SSIM among compared methods on the RESIDE and CDD datasets for all three tasks, and that preprocessing degraded images with MKoIE improves a pretrained U2Net salient object detector's MAE from 0.1491 to 0.0609 on the CDD dataset.

Load-bearing premise

The comparison assumes every method, including MKoIE, is trained and tested on the same split of the RESIDE and CDD datasets, with the test images never seen during training; the paper does not state the split or the test-image count.

Editorial extensions

If this is right

If MKoIE's reported margins hold, a single network can replace separate daytime-dehazing, low-light, and nighttime-dehazing models in a vision pipeline.
The reported 1.68-second runtime for 2560 by 1440 images suggests the architecture is light enough for near-real-time deployment on a single GPU.
Because MKoIE improves a pretrained salient object detector without retraining, it can act as a drop-in preprocessing module for downstream vision tasks.
The fusion of low-light and daytime-dehazing knowledge inside the nighttime subnode implies explicit cross-task knowledge transfer, not just shared features.

Reading between the lines

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

The attention-based fusion between the low-light and daytime-dehazing subnodes points to a general recipe: use easier degradation tasks as auxiliary supervisors for a harder composite one, which could extend to rain, snow, or underwater imaging.
MKoIE's training recipe synthesizes degraded images from a physical model; a test on real captured nighttime-haze footage would tell whether the synthetic-to-real gap is closed.
The salient-object-detection gain comes from a single pretrained detector, so the enhancer's value as a front end for other detectors remains untested.
The reported NIQE margins are small, so the no-reference quality advantage may be less robust than the PSNR and SSIM advantage; this is an inference, not the paper's claim.

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.

Desk Editor's Note

read the letter

Quick take: MKoIE is a sensible, well-engineered all-in-one image enhancer for day haze, low light, and night haze, and the paper is readable. The architecture combines a task-oriented node learning mechanism with shared and separated sub-networks, self-attention, and a multi-receptive-field module built from dilated depthwise separable convolutions. That is a reasonable incremental extension of the authors' own AoSRNet, not a new paradigm, and the paper is honest about being an engineering contribution. The ablations on TNL, MRFE, and the loss components are standard but properly executed and show each piece earns its keep. The runtime table (1.68 s on 2560×1440) is a useful practical datapoint, though calling that "real-time" is a stretch.

The soft spot is the evaluation protocol. The paper says training uses 3,183 images from RESIDE-OTS and CDD but never states how the train/test split was made, how many test images were used, or whether CDD test images are disjoint from training. That matters because CDD is the authors' own composite degradation dataset, generated from the paper's own imaging model in Section 3. The baseline numbers in Tables 1–3 are also suspiciously low compared with published results (e.g., FFA-Net at 19.47 dB, well below its usual RESIDE scores), which is what you would see if task-specific models were evaluated off-distribution on CDD with released checkpoints rather than retrained on the same data. If that is what happened, the 2–10 dB margins do not prove architectural superiority; they mostly show that MKoIE was trained on the test distribution and the baselines were not. The downstream SOD improvement in Table 7 inherits the same CDD-specific protocol, so it does not independently corroborate anything.

I am not saying the method is fake or that the authors are deliberately hiding something. The architecture is coherent and the ablations are informative. But the central empirical claim — one network beats task-specific state of the art on three tasks — is not established until the protocol is clarified and code or data splits are released. The fix is straightforward: state the split, retrain baselines on the same training set, report per-dataset numbers, and release the code.

Verdict: this deserves a serious referee, but it should come back with a request for a fully specified evaluation protocol. If the numbers hold under a fair comparison, MKoIE is a useful practical tool for autonomous driving and surveillance pipelines. As is, treat the superiority claim as conditional.

Referee Report

4 major / 5 minor

Summary. MKoIE is a multi-task image enhancer that targets daytime dehazing, low-light enhancement, and nighttime haze enhancement within a single network. The architecture combines task-oriented node learning with shared and separate encoder-decoder paths, a self-attention module, a multi-receptive-field enhancement (MRFE) block, and a hybrid L1/Charbonnier plus VGG perceptual loss. Experiments compare MKoIE with task-specific state-of-the-art methods on the RESIDE and CDD datasets, report ablations of the main modules and loss terms, and show that MKoIE preprocessing improves a pretrained U2Net salient object detector on CDD. The paper also reports an inference time of 1.68 seconds for 2560×1440 images.

Significance. If the central empirical claim were established, MKoIE would be a practically attractive single-network alternative to task-specific restoration modules in vision-driven systems, especially given its reported speed and parameter efficiency. The proposed components—task-specific node learning with adaptive fusion, a lightweight MRFE block, and a hybrid loss—are reasonable and the authors provide ablation evidence for each. However, the evaluation protocol is currently insufficient to support the superiority claim: the training/test split is undefined, the CDD benchmark is generated by the same group with the same imaging model used for training, and baseline numbers fall far below their published levels. The downstream SOD experiment is also confined to the same in-distribution CDD set without comparison to other enhancers. The architecture may be sound, but the empirical validation needs substantial reworking before the claims can be accepted.

major comments (4)

[Section 5.1.1 / Section 5.1.3] The training and testing protocol is critically under-specified. The text states only that 'the training and testing datasets include RESIDE-OTS and CDD' and that 3,183 images were used for training; it never provides the train/test split, the number of test images, or an explicit statement that CDD test images are disjoint from the training set. Likewise, Section 5.1.3 says only that all baselines use 'the source code published by the authors,' leaving it unknown whether baselines were retrained on the same training distribution or applied with their released checkpoints. Without a precise protocol, the large margins in Tables 1–3 cannot be attributed to architectural superiority.
[Section 5.1.1 + Section 3 + ref [49]] The evaluation is potentially circular. CDD (ref [49]) is a composite-degradation dataset produced by the authors' own research group, and Section 5.1.1 says degraded images are synthesized 'based on the imaging model in Sec. 3.' MKoIE is thus trained and tested on degradations generated from the same physical model that also defines the proposed network's target distribution. This gives MKoIE an in-distribution advantage over task-specific baselines that were not trained on this composite degradation model. The authors need to evaluate on independent public benchmarks (e.g., RESIDE SOTS, LOL, and standard nighttime-haze sets) and report per-dataset results for both MKoIE and all baselines under a single, clearly described protocol.
[Tables 1–3] The baseline results are far below published state-of-the-art numbers, which is a strong signal that the comparison is not being run under the baselines' native conditions. For example, FFA-Net reports 19.465 dB PSNR in Table 1, whereas its published performance on RESIDE SOTS is around 36 dB; TaylorFormerV2 at 23.827 dB is likewise substantially below its reported benchmark. If the baselines were evaluated on CDD or on a mixed distribution without retraining, the reported margins do not establish that MKoIE outperforms these methods on their own tasks. The authors must specify whether each baseline was retrained on the same training set, and if not, why an off-the-shelf checkpoint comparison is appropriate for a fairness claim.
[Section 5.5 / Table 7] The downstream SOD experiment only compares original CDD images against MKoIE-enhanced images using a pretrained U2Net. It does not compare MKoIE with any other enhancement method, so it does not support the broader claim that MKoIE is the preferable enhancer for intelligent-vision pipelines. Moreover, because CDD is the same dataset used for training MKoIE, the improvement may reflect in-distribution adaptation rather than general improvement of SOD on genuinely unseen degraded imagery. A comparison against one or two strong task-specific enhancers on an independent SOD benchmark would be needed.

minor comments (5)

[Section 5.6 / Table 8] The runtime comparison is not fully controlled: IENHC and MRP are in C++ while ECNHI and MKoIE are in PyTorch, and the GPU model used for testing is not stated (Section 5.1.4 lists the training GPUs but not the inference setup). The claim of 'consistent hardware setup' needs explicit specification, and mixed-language comparisons should be interpreted with caution.
[Throughout] Typos and minor language issues: Section 4.5 'various erent types' should read 'various different types'; Section 5.2.3 'MKoIE cexhibits' should read 'MKoIE exhibits'; Section 5.2.1 refers to 'TaylorFormer' while the table lists 'TaylorFormerV2'; and reference [42] has a duplicated title fragment.
[Section 5.1.4] The optimizer is described as 'adaptive moment estimation optimizer'—use the standard name Adam. Also, the learning-rate schedule (decay at epochs 60, 120, 180) is reported only for MKoIE; the same training details should be provided for baselines if they were retrained.
[Section 5.1.1] The term 'RESIDE-OTS' and the table header 'RESIDE' are used inconsistently; it should be clear whether the test set is from RESIDE-OTS or the RESIDE SOTS benchmark, since these have different properties.
[Reproducibility] No code or pretrained models are released for MKoIE, and the paper does not state an intention to release them. Given the protocol ambiguities, providing code and trained weights would materially help reviewers and readers verify the results.

Circularity Check

3 steps flagged · score 7.0 of 10

Central benchmark reduces to in-distribution evaluation: MKoIE is trained and tested on the authors' own CDD dataset synthesized from their Eq. (1) imaging model, with no disclosed split; the downstream SOD validation is also on the same training distribution.

fitted input called prediction [Section 5.1.1 and Tables 1-3]
"The training and testing datasets include realistic single image dehazing (RESIDE)-OTS [48], which incorporates depth information, and the composite degradation dataset (CDD) [49]. Based on the imaging model in Sec. 3, we extract atmospheric light values according to real conditions to synthesize more realistic degraded images [22]."

The central superiority claims in Tables 1-3 are reported on 'RESIDE and CDD datasets'. As described, CDD is both a training dataset and a testing dataset, and the degraded images are synthesized from the paper's own Eq. (1) imaging model. No train/test split or test-image count is specified, so the reported PSNR/SSIM margins are measured on the same named dataset and the same degradation pipeline used to train MKoIE. The benchmark 'prediction' is therefore an in-distribution fit rather than an independent, held-out evaluation of the claimed advantage over task-specific baselines.
self citation load bearing [References [22] and [49], used in Section 5.1.1]
"[22] Y. Lu, D. Yang, Y. Gao, R. W. Liu, J. Liu, Y. Guo, Aosrnet: All-in-one scene recovery networks via multi-knowledge integration, Knowledge-Based Syst. 294 (2024) 111786. [49] Y. Guo, Y. Gao, Y. Lu, H. Zhu, R. W. Liu, S. He, Onerestore: A universal restoration framework for composite degradation, in: Proc. ECCV, Springer, 2025, pp. 255-272."

The evaluation's test distribution is CDD, cited to [49], and the degradation synthesis recipe is cited to [22]. Both references include MKoIE co-authors Yuxu Lu and Dong Yang. These self-citations are load-bearing because the entire quantitative comparison—ID, LLIE, NHIE, and the downstream SOD experiment—uses CDD or the CDD-corruption model as the benchmark. The authors are evaluating on their own composite-degradation dataset and citing their own prior work to establish it as the benchmark, with no independent external test set such as SOTS with published baseline scores.

1 more flagged steps

fitted input called prediction [Section 5.5, Table 7]
"We utilized U2Net [52] and show its performance on CDD dataset. Notably, we directly applied the pretrained U2Net model without further training, achieving a marked enhancement in SOD through MKoIE preprocessing."

The downstream SOD validation is performed on CDD, which is one of MKoIE's training datasets (Section 5.1.1). The reported improvement (MAE from 0.1491 to 0.0609) is therefore measured on the distribution MKoIE was optimized to restore. It is a fitted-input demonstration rather than independent evidence that MKoIE improves SOD on unseen or real-world degraded imagery, and it inherits the same in-distribution protocol as the restoration tables.

full rationale

The architectural derivation itself is not internally circular: the TNL, SA, MRFE, and hybrid loss are defined as standard network components, and the ablations are ordinary training comparisons. The circularity is in the evidence for the central claim that one network surpasses task-specific methods. As reported, MKoIE is trained and evaluated on CDD, a dataset produced by the authors' own OneRestore pipeline, with degraded images synthesized from the paper's own Eq. (1) model, and no train/test split is disclosed. The downstream SOD test in Section 5.5 is also on CDD. These are not external checks; they measure performance on the authors' own degradation model and dataset. The self-citations to [22] and [49] are load-bearing because they supply both the corruption synthesis and the benchmark distribution. I assign 7 rather than 8-10 because the method has independent architectural content and could plausibly transfer, but the superiority claim as presented is predominantly supported by self-generated, in-distribution evaluation rather than by independent validation.

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

The central claim inherits three families of assumptions: the compound imaging model in Eq. 1 is treated as an adequate model of real degradation; the CDD/RESIDE datasets are treated as representative and are split in an undisclosed way; and the perceptual loss and SOD probe are assumed beneficial. The free parameters are mostly hand-set hyperparameters (loss weights, dilation rates, learning rate schedule) plus one learned fusion weight alpha. No invented physical entities are introduced.

free parameters (4)

Loss weights w1 and w2 = 0.8 and 0.2
Set manually in Section 4.6.3 after experiments; the paper says this configuration 'demonstrates good balance' rather than being derived.
MRFE dilation rates = 1, 3, 5
Hand-chosen in Section 4.4 to cover 3x3, 7x7, and 11x11 receptive fields; no sensitivity analysis is reported.
Feature fusion weight alpha = learned via sigmoid
Learnable scalar in Eq. (6) balancing LLIE-subnode and ID-subnode features for the NHIE task; fitted during training.
Learning rate schedule epochs = 1e-3, decay x10 at 60/120/180
Training hyperparameter in Section 5.1.4; chosen without sensitivity analysis.

assumptions (4)

domain assumption The compound imaging model I(p) = J(p)L(p)t(p) + A(1 - t(p)) + N(p) accurately describes daytime haze, low light, and nighttime haze degradation.
Used in Section 3 and to synthesize degraded training images in Section 5.1.1; not validated against real physical measurements.
domain assumption CDD and the synthesized RESIDE variants are representative of real-world degraded images.
The quantitative claims are evaluated on these datasets; real-world evaluation in Section 5.3 is qualitative only.
domain assumption VGG-16 features provide a useful perceptual loss for all three restoration tasks.
Perceptual loss in Eq. (10) uses pretrained VGG-16 layers; the choice of layers conv1_2, conv2_2, conv3_3 is asserted without ablation.
domain assumption Pretrained U2Net is a valid and unbiased probe for downstream SOD improvement.
Section 5.5 uses U2Net without retraining; no analysis of whether the improvement is specific to U2Net or generalizes to other detectors.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Knowledge-oriented Nighttime Haze Imaging Enhancer for Vision-driven Intelligent Systems." pith.science (2026). https://pith.science/paper/LCZZXME3

@misc{pith2026250207351,
  author       = {Pith},
  title        = {Pith review of: Multi-Knowledge-oriented Nighttime Haze Imaging Enhancer for Vision-driven Intelligent Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCZZXME3}},
  note         = {Machine review of arXiv:2502.07351}
}

read the original abstract

Salient object detection (SOD) plays a critical role in Intelligent Imaging, facilitating the detection and segmentation of key visual elements in an image. However, adverse imaging conditions such as haze during the day, low light, and haze at night severely degrade image quality and hinder reliable object detection in real-world scenarios. To address these challenges, we propose a multi-knowledge-oriented nighttime haze imaging enhancer (MKoIE), which integrates three tasks: daytime dehazing, low-light enhancement, and nighttime dehazing. The MKoIE incorporates two key innovative components: First, the network employs a task-oriented node learning mechanism to handle three specific degradation types: day-time haze, low light, and night-time haze conditions, with an embedded self-attention module enhancing its performance in nighttime imaging. In addition, multi-receptive field enhancement module that efficiently extracts multi-scale features through three parallel depthwise separable convolution branches with different dilation rates, capturing comprehensive spatial information with minimal computational overhead to meet the requirements of real-time imaging deployment. To ensure optimal image reconstruction quality and visual characteristics, we suggest a hybrid loss function. Extensive experiments on different types of weather/imaging conditions illustrate that MKoIE surpasses existing methods, enhancing the reliability, accuracy, and operational efficiency of intelligent imaging.

Figures

Figures reproduced from arXiv: 2502.07351 by the authors.

**Figure 2.** Overall architecture of the multi-knowledge-oriented nighttime haze image en [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** The pipelines of residual learning block (RLB) and multi-receptive field enhance [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: The pipeline of self attention-guided node learning module that includes self [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

**Figure 5.** Figure 5: Visual comparisons with the SOTA ID, LLIE, and NHIE methods on the RE [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

**Figure 6.** Figure 6: Qualitative comparisons with the SOTA NHIE methods on the CDD datasets. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

**Figure 7.** Figure 7: Visual comparisons on synthetic degradation scene. Zooming-in is recommended [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

**Figure 8.** Figure 8: Visual comparisons on real-world degradation scenes. Zooming-in is recom [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]

**Figure 9.** Figure 9: Illustration of MKoIE without TNL components. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]

**Figure 10.** Figure 10: Comparisons of U2Net-based SOD results for visually-degraded images and [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 45 canonical work pages

[49]

Y. Guo, Y. Gao, Y. Lu, H. Zhu, R. W. Liu, S. He, Onerestore: A universal restoration framework for composite degradation, in: Proc. ECCV, Springer, 2025, pp. 255–272

work page 2025
[1]

W. Wang, Q. Lai, H. Fu, J. Shen, H. Ling, R. Yang, Salient object detection in the deep learning era: An in-depth survey, IEEE Trans. Pattern Anal. Mach. Intell. 44 (6) (2021) 3239–3259

work page 2021
[2]

T. Feng, C. Wang, X. Chen, H. Fan, K. Zeng, Z. Li, Urnet: A u-net based residual network for image dehazing, Applied Soft Computing 102 (2021) 106884

work page 2021
[3]

M. C.V., D. R. M., S. D., On generalized sugeno’s class generator and parametrized intuitionistic fuzzy approach for enhancing low-light images, Applied Soft Computing 172 (2025) 112865

work page 2025
[4]

Hummel, Image enhancement by histogram transformation, Comput

R. Hummel, Image enhancement by histogram transformation, Comput. Graphics and Image Proce. 6 (2) (1977) 184–195

work page 1977
[5]

E. H. Land, The retinex theory of color vision, Sci. Am. Mind 237 (6) (1977) 108–129

work page 1977
[6]

X. Guo, Y. Li, H. Ling, Lime: Low-light image enhancement via illumination map estimation, IEEE Trans. Image Process. 26 (2) (2016) 982–993

work page 2016
[7]

S. Hao, X. Han, Y. Guo, X. Xu, M. Wang, Low-light image enhancement with semi- decoupled decomposition, IEEE Trans. Multimedia 22 (12) (2020) 3025–3038

work page 2020

Show all 52 references

[8]

K. He, J. Sun, X. Tang, Single image haze removal using dark channel prior, IEEE Trans. Pattern Anal. Mach. Intell. 33 (12) (2010) 2341–2353

2010
[9]

J. Liu, R. W. Liu, J. Sun, T. Zeng, Rank-one prior: Real-time scene recovery, IEEE Trans. Pattern Anal. Mach. Intell. 45 (7) (2022) 8845–8860

2022
[10]

Zhang, Y

J. Zhang, Y. Cao, Z. Wang, Nighttime haze removal based on a new imaging model, in: Proc. IEEE ICIP, 2014, pp. 4557–4561

2014
[11]

Y. Li, R. T. Tan, M. S. Brown, Nighttime haze removal with glow and multiple light colors, in: Proc. IEEE ICCV, 2015, pp. 226–234

2015
[12]

Zhang, Y

J. Zhang, Y. Cao, S. Fang, Y. Kang, W. C. Chang, Fast haze removal for nighttime image using maximum reflectance prior, in: Proc. IEEE CVPR, 2017, pp. 7418–7426. 25

2017
[13]

S. Lin, F. Tang, W. Dong, X. Pan, C. Xu, Smnet: Synchronous multi-scale low light enhancement network with local and global concern, IEEE Trans. Multimedia 25 (2023) 9506–9517

2023
[14]

X. Qin, Z. Wang, Y. Bai, X. Xie, H. Jia, Ffa-net: Feature fusion attention network for single image dehazing, in: Proc. AAAI, Vol. 34, 2020, pp. 11908–11915

2020
[15]

X. Liu, Z. Shi, Z. Wu, J. Chen, G. Zhai, Griddehazenet+: An enhanced multi-scale network with intra-task knowledge transfer for single image dehazing, IEEE Trans. Intell. Transp. Syst. 24 (1) (2022) 870–884

2022
[16]

Y. Song, Z. He, H. Qian, X. Du, Vision transformers for single image dehazing, IEEE Trans. Image Process. 32 (2023) 1927–1941

2023
[17]

S. Yin, S. Hu, Y. Wang, Y.-H. Yang, High-order adams network (hian) for image dehazing, Applied Soft Computing 139 (2023) 110204

2023
[18]

Zhang, X

S. Zhang, X. Zhang, S. Wan, W. Ren, L. Zhao, L. Shen, Generative adversarial and self-supervised dehazing network, IEEE Trans. Ind. Inf. 20 (3) (2023) 4187–4197

2023
[19]

Zheng, X

Z. Zheng, X. Jia, 4k-haze: A dehazing benchmark with 4k resolution hazy and haze- free images, arXiv preprint arXiv:2303.15848 (2023)

2023 arXiv
[20]

J. Hou, Z. Zhu, J. Hou, H. Liu, H. Zeng, H. Yuan, Global structure-aware diffusion process for low-light image enhancement, Adv. Neural Inform. Process. Syst. 36 (2024) 79734–79747

2024
[21]

Liang, Y

H. Liang, Y. Yang, J. Jing, Latent space diffusion model for image dehazing, Applied Soft Computing 180 (2025) 113322

2025
[22]

Y. Lu, D. Yang, Y. Gao, R. W. Liu, J. Liu, Y. Guo, Aosrnet: All-in-one scene recovery networks via multi-knowledge integration, Knowledge-Based Syst. 294 (2024) 111786

2024
[23]

Y. Jin, B. Lin, W. Yan, Y. Yuan, W. Ye, R. T. Tan, Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, in: Proc. ACM MM, 2023, pp. 2446–2457

2023
[24]

H. Yin, P. Yang, Multi-stage progressive single image dehazing network with feature physics model, IEEE Trans. Instrum. Meas. 73 (2024) 1–12

2024
[25]

B. Cai, X. Xu, K. Jia, C. Qing, D. Tao, Dehazenet: An end-to-end system for single image haze removal, IEEE Trans. Image Process. 25 (11) (2016) 5187–5198

2016
[26]

B. Li, X. Peng, Z. Wang, J. Xu, D. Feng, Aod-net: All-in-one dehazing network, in: Proc. IEEE ICCV, 2017, pp. 4770–4778

2017
[27]

W. Xu, D. Yang, Y. Gao, Y. Lu, J. Zhang, Y. Guo, Mvksr: Multi-view knowledge- guided scene recovery for hazy and rainy degradation, IEEE Trans. Instrum. Meas. (2024). 26

2024
[28]

X. Gong, Z. Zheng, H. Du, Tsnet: a two-stage network for image dehazing with multi-scale fusion and adaptive learning, Signal, Image Video Process. 18 (10) (2024) 7119–7130

2024
[29]

C. Guo, C. Li, J. Guo, C. C. Loy, J. Hou, S. Kwong, R. Cong, Zero-reference deep curve estimation for low-light image enhancement, in: Proc. IEEE ICCV, 2020, pp. 1780–1789

2020
[30]

W. Wu, J. Weng, P. Zhang, X. Wang, W. Yang, J. Jiang, Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement, in: Proc. IEEE ICCV, 2022, pp. 5901–5910

2022
[31]

Y. Liao, Z. Su, X. Liang, B. Qiu, Hdp-net: Haze density prediction network for nighttime dehazing, in: Proc. PCM, Springer, 2018, pp. 469–480

2018
[32]

X. Cong, J. Gui, J. Zhang, J. Hou, H. Shen, A semi-supervised nighttime dehazing baseline with spatial-frequency aware and realistic brightness constraint, in: Proc. IEEE CVPR, 2024, pp. 2631–2640

2024
[33]

Zhang, Y

J. Zhang, Y. Cao, Z. Wang, Nighttime haze removal with illumination correction, arXiv preprint arXiv:1606.01460 (2016)

2016 arXiv
[34]

B. Koo, G. Kim, Nighttime haze removal with glow decomposition using gan, in: Proc. ACPR, Springer, 2020, pp. 807–820

2020
[35]

Zhang, J

Y. Zhang, J. Zhang, X. Guo, Kindling the darkness: A practical low-light image enhancer, in: Proc. ACM MM, 2019, pp. 1632–1640

2019
[36]

J. Yu, Q. Liao, Fast single image fog removal using edge-preserving smoothing, in: Proc. IEEE ICASSP, IEEE, 2011, pp. 1245–1248

2011
[37]

Fattal, Dehazing using color-lines, ACM Trans

R. Fattal, Dehazing using color-lines, ACM Trans. Graphics 34 (1) (2014) 1–14

2014
[38]

S. M. Pizer, E. P. Amburn, J. D. Austin, R. Cromartie, A. Geselowitz, T. Greer, B. ter Haar Romeny, J. B. Zimmerman, K. Zuiderveld, Adaptive histogram equalization and its variations, Computer vision, graphics, and image processing 39 (3) (1987) 355–368

1987
[39]

Y. Cai, H. Bian, J. Lin, H. Wang, R. Timofte, Y. Zhang, Retinexformer: One-stage retinex-based transformer for low-light image enhancement, in: Proc. IEEE ICCV, 2023, pp. 12504–12513

2023
[40]

Pei, T.-Y

S.-C. Pei, T.-Y. Lee, Nighttime haze removal using color transfer pre-processing and dark channel prior, in: Proc. IEEE ICIP, IEEE, 2012, pp. 957–960

2012
[41]

X. Liu, H. Li, C. Zhu, Joint contrast enhancement and exposure fusion for real-world image dehazing, IEEE Trans. Multimedia 24 (2021) 3934–3946. 27

2021
[42]

G. R. dense transformer with grid structure for image restoration in adverse weather conditions, Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions, International Journal of Computer Vision (2024) 1–23

2024
[43]

Z. Jin, Y. Qiu, K. Zhang, H. Li, W. Luo, Mb-taylorformer v2: Improved multi-branch linear transformer expanded by taylor formula for image restoration, TPAMI (2025)

2025
[44]

C. Liu, F. Wu, X. Wang, Efinet: Restoration for low-light images via enhancement- fusion iterative network, IEEE Trans. Circuits Syst. Video Technol. 32 (12) (2022) 8486–8499

2022
[45]

C. He, C. Fang, Y. Zhang, K. Li, L. Tang, C. You, F. Xiao, Z. Guo, X. Li, Reti-diff: Illumination degradation image restoration with retinex-based latent diffusion model (2023). arXiv:2311.11638

2023 arXiv
[46]

Chan, W.-C

C.-Y. Chan, W.-C. Siu, Y.-H. Chan, H. Anthony Chan, Anlightendiff: Anchoring diffusion probabilistic model on low light image enhancement, IEEE Transactions on Image Processing 33 (2024) 6324–6339

2024
[47]

Zhang, Y

J. Zhang, Y. Cao, Z.-J. Zha, D. Tao, Nighttime dehazing with a synthetic benchmark, in: Proc. ACM MM, 2020

2020
[48]

B. Li, W. Ren, D. Fu, D. Tao, D. Feng, W. Zeng, Z. Wang, Benchmarking single-image dehazing and beyond, IEEE Trans. Image Process. 28 (1) (2019) 492–505

2019
[50]

Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli, Image quality assessment: From error visibility to structural similarity, IEEE Trans. Image Process. 13 (4) (2004) 600– 612

2004
[51]

completely blind

A. Mittal, R. Soundararajan, A. C. Bovik, Making a “completely blind” image quality analyzer, IEEE Signal Process Lett. 20 (3) (2012) 209–212

2012
[52]

X. Qin, Z. Zhang, C. Huang, M. Dehghan, O. R. Zaiane, M. Jagersand, U2-net: Going deeper with nested u-structure for salient object detection, Pattern Recognit. 106 (2020) 107404. 28

2020

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [49]

Y. Guo, Y. Gao, Y. Lu, H. Zhu, R. W. Liu, S. He, Onerestore: A universal restoration framework for composite degradation, in: Proc. ECCV, Springer, 2025, pp. 255–272

work page 2025

[2] [1]

W. Wang, Q. Lai, H. Fu, J. Shen, H. Ling, R. Yang, Salient object detection in the deep learning era: An in-depth survey, IEEE Trans. Pattern Anal. Mach. Intell. 44 (6) (2021) 3239–3259

work page 2021

[3] [2]

T. Feng, C. Wang, X. Chen, H. Fan, K. Zeng, Z. Li, Urnet: A u-net based residual network for image dehazing, Applied Soft Computing 102 (2021) 106884

work page 2021

[4] [3]

M. C.V., D. R. M., S. D., On generalized sugeno’s class generator and parametrized intuitionistic fuzzy approach for enhancing low-light images, Applied Soft Computing 172 (2025) 112865

work page 2025

[5] [4]

Hummel, Image enhancement by histogram transformation, Comput

R. Hummel, Image enhancement by histogram transformation, Comput. Graphics and Image Proce. 6 (2) (1977) 184–195

work page 1977

[6] [5]

E. H. Land, The retinex theory of color vision, Sci. Am. Mind 237 (6) (1977) 108–129

work page 1977

[7] [6]

X. Guo, Y. Li, H. Ling, Lime: Low-light image enhancement via illumination map estimation, IEEE Trans. Image Process. 26 (2) (2016) 982–993

work page 2016

[8] [7]

S. Hao, X. Han, Y. Guo, X. Xu, M. Wang, Low-light image enhancement with semi- decoupled decomposition, IEEE Trans. Multimedia 22 (12) (2020) 3025–3038

work page 2020