Overcoming Attention Drift: Homogeneity-Heterogeneity Guided Feature Aggregation for Low-Light Remote Sensing Image Enhancement

Mingyang Ma; Shaohui Mei; Xingxing Yang; Yaozi Zhong

REVIEW 5 major objections 4 minor 50 references

Overcoming Attention Drift: Homogeneity-Heterogeneity Guided Feature Aggregation for Low-Light Remote Sensing Image Enhancement

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adding semantic and geometric terms to attention logits prevents low-light remote sensing features from merging across physical boundaries.

desk verdict A genuinely different way to inject semantic and depth priors into attention logits, with solid zero-shot experiments, but the load-bearing assumption that DINOv3/DA3 priors survive extreme low-light input is deferred to a missing appendix. read the letter →

arxiv 2608.05843 v1 pith:4DVCCTXK submitted 2026-08-06 cs.CV

Yaozi Zhong , Xingxing Yang , Shaohui Mei , Mingyang Ma This is my paper · ORCID

classification cs.CV

keywords low-lightremotesensingimageenhancementattentiondriftfeatureaggregationfoundationmodelpriorssemantichomogeneitygeometricheterogeneityobjectdetection

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

Under extreme low-light conditions, self-attention in restoration networks aggregates features based on corrupted statistics, so pixels from different physical objects get merged and continuous regions get split. This paper claims that this "attention drift" is the bottleneck, and that it can be removed by making attention obey external physical constraints rather than pure data. The proposed HALO framework takes semantic features from a frozen illumination-invariant model as a positive bias that pulls tokens within the same region together, and depth-boundary features from a frozen monocular depth model as a negative penalty that blocks aggregation across edges. The authors add both biases directly to the attention logits through a module called H2CAM, and report results that outperform the compared methods on eight synthetic and real-world remote sensing benchmarks, with sharper boundaries, better color fidelity, and improved downstream object detection. If correct, the idea shows a concrete way to keep foundation-model priors from being diluted by noisy observations.

What carries the argument

The load-bearing mechanism is the Homogeneity-Heterogeneity Cooperative Attention Module (H2CAM), a bottleneck attention layer whose logits are $L = QK^\top/\sqrt{d} + \lambda_s B_{\mathrm{sem}} + \lambda_g B_{\mathrm{geo}}$. $B_{\mathrm{sem}}$ is the L2-normalized cosine similarity of frozen semantic features, injected as a positive bias to pull tokens toward their semantic cluster; $B_{\mathrm{geo}}$ is $-\alpha|g_i - g_j|$ over a depth-boundary descriptor, injected as a negative penalty that acts as a wall at physical edges. The gating scalars $\lambda_s$ and $\lambda_g$ are learnable and non-negative, and the module sits at the bottleneck of the final cascaded encoder-decoder stage, where the receptive field is largest, so the priors constrain aggregation at the point of maximum semantic and geometric reasoning.

What would settle it

Measure the agreement between the depth model's predicted boundaries and true object edges on a low-light remote sensing test set with known geometry as noise increases; if the boundary maps used for $B_{\mathrm{geo}}$ disagree with real edges at the noise levels where HALO still reports high PSNR, the hard constraints are enforcing the wrong walls. A minimal version is to take one iSAID-dark image, add noise, and check whether the $|g_i - g_j|$ wall between a building and its shadow coincides with the actual object edge.

Watch

Extended reading notes

Core claim

The central claim is that unconstrained self-attention fails in extreme low-light remote sensing because the affinity matrix $A_{i,j} \propto \exp(q_i k_j^\top/\sqrt{d})$ is computed from features corrupted by illumination and noise, so the true semantic correlation is overwhelmed by two failure modes: Type I cross-boundary confusion, where noise makes physically separate tokens look similar, and Type II coplanar ambiguity, where different materials collapse to the same dark response while homogeneous regions fragment. HALO treats enhancement as a guided feature aggregation problem and rewrites the attention logits as $L = QK^\top/\sqrt{d} + \lambda_s B_{\mathrm{sem}} + \lambda_g B_{\mathrm{geo}}$, where $B_{\mathrm{sem}}$ is a positive cosine-similarity bias from an illumination-invariant semantic prior and $B_{\mathrm{geo}}$ is a negative absolute-difference penalty from a pseudo-3D depth prior. With these two constraints inside the attention computation rather than concatenated as auxiliary inputs, the paper argues that aggregation is bounded by physical scene structure, eliminating both drift types and producing sharper boundaries, faithful colors, and better downstream detection.

Load-bearing premise

The method assumes that the semantic and depth priors extracted from the dark, noisy input are accurate enough to define true region and boundary constraints; the paper defers the check of that assumption to an appendix section that is not present in the reviewed text.

Editorial extensions

If this is right

Attention-based low-light enhancers that ignore physical boundaries will keep failing in exactly the regimes HALO improves: the paper's ablations show that simply concatenating both priors into the features without H2CAM drops PSNR to 25.315 dB, while the full guided aggregation reaches 26.527 dB on iSAID-dark.
The priors carry the generalization: a model trained only on iSAID-dark transfers zero-shot to iSAID-dark (high-pixel), UCM-RSLL, DarkRS, and U3D, without retraining or fine-tuning.
Downstream Earth-observation tasks inherit the benefit: on low-light DOTA-v1.0, HALO-enhanced images give a pre-trained detector an F1-score of 0.790 and mAP@0.5 of 0.732, narrowing the gap to the clean-image upper bound.
Performance degrades gradually rather than collapsing under increasing Gaussian noise, with PSNR falling from 21.893 dB at $\sigma=0$ to 18.921 dB at $\sigma=20$, consistent with the claim that the two priors are photometrically invariant.

Reading between the lines

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

Because $B_{\mathrm{geo}}$ is defined from any reliable discontinuity map, the same additive logit-bias recipe could transfer to other restoration problems where edges are known a priori, such as depth-aware dehazing or specular highlight removal; the paper itself does not make this claim.
The paper defers validation of prior reliability to an appendix, so the ceiling of the method is set by the frozen priors' accuracy under darkness and noise, not by the attention module alone.
The learned gating scalars $\lambda_s$ and $\lambda_g$ could be read as a per-image diagnostic of where attention drift is most severe, providing a cheap confidence map for downstream tasks; this is a use the paper leaves implicit.
Since roughly half of the reported 3966.64 ms inference time on U3D comes from extracting the two frozen priors, a practical next step would distill the priors into a lightweight single-network estimator; the paper does not address this.

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

The actual novelty here is real. Prior-guided low-light methods mostly concatenate semantic or depth features and hope the network uses them. HALO instead adds a semantic homogeneity bias and a geometric heterogeneity penalty directly into the attention logits (Eq. 10), and the ablations show this beats both single-prior and naive double-prior concatenation. That is a concrete, testable mechanism, and the zero-shot evaluations on iSAID-dark high-pixel, UCM-RSLL, DarkRS, U3D, and DOTA give the central claim independent grounding. I also give the authors credit for reporting second place on LOLv1 PSNR and for breaking out the inference time overhead of the foundation models. The paper is not overselling itself everywhere.

That said, the main soft spot is the one the stress test flags, and it is real. Equation (6) computes the DINOv3 and DA3 priors from the same severely degraded input I, before any restoration. The whole method depends on those priors being accurate under extreme low light and noise, yet the only support is a sentence pointing to Appendix Sec. IV, which is not in the reviewed text. The noise robustness study in Table VII only shows the final PSNR degrading gracefully; it does not show that the priors themselves remain faithful. If the priors are corrupted, B_sem will group wrong regions and B_geo will penalize legitimate aggregation, and because the biases sit inside the final-stage attention, the error propagates directly into the output. The paper calls this an "impenetrable mathematical wall," but lambda_s and lambda_g are trainable scalars, so there is no guarantee the wall is even up. This is a conditional, not fatal, problem, but it is the first thing a referee should push on.

Other issues are smaller. Equation (2) is presented as a formalization but is really a heuristic decomposition; nobody should mistake it for a derivation. There are no error bars or significance tests on the tables. The code link is a placeholder, and the appendices that supposedly contain training details and prior reliability validation are absent. The state-of-the-art claim is mostly true but not universal, and the paper acknowledges that.

Bottom line: the mechanism is promising, the experiments are broad and mostly consistent, and the zero-shot evidence is a real strength. The missing appendices and unvalidated prior-reliability assumption are exactly what peer review should fix. I would send it out, and I would ask for the code, the appendices, and a direct comparison of priors on clean versus degraded inputs before accepting.

For our purposes: worth reading for anyone working on low-light remote sensing or prior-guided restoration, but wait until the appendix and code are actually available before building on it.

Referee Report

5 major / 4 minor

Summary. The paper proposes HALO, a low-light remote sensing image enhancement method that combines a semantic prior from frozen DINOv3 features and a geometric prior from Depth Anything 3 depth maps. These priors are converted into additive bias terms in attention logits: a positive homogeneity bias B_sem based on cosine similarity of semantic tokens and a negative heterogeneity penalty B_geo based on absolute depth differences. The resulting Homogeneity-Heterogeneity Cooperative Attention Module (H2CAM) is inserted at the bottleneck of the final restoration stage. The paper introduces a conceptual 'Attention Drift' failure mode with Type I (cross-boundary confusion) and Type II (coplanar ambiguity) errors, and claims that the dual priors provide deterministic physical constraints that resolve this bottleneck. Evaluation is broad: full-reference benchmarks (iSAID-dark, iSAID-dark high-pixel, UCM-RSLL, LOLv1, LSRW-Nikon), no-reference real-world benchmarks (DarkRS, U3D), and downstream detection on DOTA-v1.0, including zero-shot cross-dataset evaluation for several benchmarks.

Significance. If the reported results hold, HALO is a practically useful contribution: the gains on iSAID-dark, iSAID-dark high-pixel, UCM-RSLL, and DOTA detection are consistent, the zero-shot held-out evaluation is a genuine strength, and the paper states that code will be released. The central design idea, translating low-level priors into explicit attention biases rather than concatenating them, is plausible and well motivated. However, the theoretical formalization is asserted rather than derived, and the validity of the frozen foundation-model priors on severely degraded inputs—the load-bearing assumption of the mechanism—is deferred to an appendix that is not present in the reviewed text. Because the biases are added inside the attention logits, any error in those priors is directly inherited by the restored image, so this missing validation is not a minor omission.

major comments (5)

[§III-A, Eq. (2)] Equation (2) is presented as a mathematical formalization of attention drift, but no derivation is given. The symbols q_J, k_J, l_i, and n_i are not defined precisely, and it is not explained how the linear degradation model I(x) = J(x)⊙L(x)+N(x) translates into the feature-domain expression after query/key projection. Without this derivation, the split into Type I and Type II errors is a heuristic narrative rather than a proven decomposition. Please either derive Eq. (2) from the attention computation with explicit notation, or state clearly that it is a motivating approximation, and characterize the regime in which the claimed l→0 and noise-dominated limits hold.
[§III-C and Appendix Sec. IV] The central mechanism relies on DINOv3 semantic features and Depth Anything 3 depth maps computed from the severely degraded input I, before any restoration. The manuscript asserts that the DA3 prior is reliable and points to Appendix Sec. IV for validation, but that appendix is absent from the submitted text. Table VII only shows that the whole HALO pipeline degrades gracefully under added Gaussian noise; it does not measure whether the semantic clusters or depth boundaries themselves remain accurate on dark, noisy inputs. Since B_sem and B_geo are applied inside the attention logits, incorrect priors would actively enforce wrong grouping. Please include the promised prior-consistency and boundary-overlap analysis, and add failure-case discussion for inputs where DINOv3/DA3 are known to be uncertain.
[§III-D, Eqs. (9)–(10)] The text describes B_geo as an 'impenetrable mathematical wall,' but this is not guaranteed by the formulation. The gating scalars λ_s and λ_g are trainable and only constrained to be nonnegative, so λ_g could become arbitrarily small or zero during training, effectively disabling the geometric penalty. A true hard wall would require a lower bound on λ_g or a non-learnable threshold, or at minimum a report of the learned λ values. In addition, |g_i - g_j| is not normalized relative to the scale of QK^T/√d, so the relative strength of the penalty is unclear across windows and datasets. Please clarify the actual range of the penalty and the learned gating behavior.
[§III-B and §III-D] H2CAM is introduced only at the bottleneck of the final restoration stage, yet the framework description in Eq. (5) and Fig. 3(a) claims that the geometric prior acts as a persistent structural anchor throughout all stages and that the semantic prior is injected at the bottleneck of each stage. No mechanism is specified for how G0 and F_sem constrain attention in stages 1 and 2, or whether those stages still use unconstrained self-attention. If earlier stages use unconstrained attention, the claim that HALO fundamentally resolves attention drift is overstated, because cross-boundary leakage would already occur before the final H2CAM. Please specify the guidance mechanism at every stage, or explicitly state that H2CAM is applied only at the final bottleneck and explain why this is sufficient.
[Table III and Table II] The LSRW-Nikon results do not fully support the claimed structural superiority: HALO's SSIM of 0.470 is below DFFN (0.481), LersGAN (0.572), and SPJFNet (0.503), even though the text emphasizes boundary sharpness and structural preservation. The paper reports only LPIPS, MAE, and ΔE00 as evidence, and does not discuss the SSIM gap. Similarly, in Table II on DarkRS, HALO's NIQE of 4.909 is worse than DFFN's 3.618 and CIDNet's 3.398, and this is not discussed. Please reconcile these numbers with the claims, and add repeated-run variability or significance tests so that readers can judge whether the headline improvements are robust.

minor comments (4)

[Appendix references] The text repeatedly defers to Appendix Secs. I–V for dataset construction, training hyperparameters, objective functions, and prior validation, but no appendix is included in the reviewed manuscript. The paper is therefore not self-contained, and these details must be supplied.
[Title page] The code link is given as 'HALO' without a URL; please provide a working repository address.
[Tables I–IV] All quantitative tables report single values without error bars or standard deviations. Given that several margins are small (for example, 26.527 dB vs. 25.826 dB on iSAID-dark), reporting mean ± std over multiple runs would materially strengthen the claims.
[Table VII] The noise-robustness study reports only HALO's own performance under increasing σ; without the same protocol applied to baseline methods, the claim that HALO 'neutralizes noise-induced structural blurring' better than existing methods is not comparatively supported.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HALO's gains are grounded in held-out zero-shot benchmarks and external frozen priors, not in fitted inputs or self-citation.

full rationale

The paper's derivation chain is an architectural proposal rather than a closed-form derivation from first principles. The attention-drift formulation in Eq. (2) is a descriptive model of degradation, not an equation whose output is presupposed by its inputs. The two bias terms in Eq. (10) are explicit functions of external frozen foundation-model outputs (DINOv3 features and Depth Anything 3 depth), and their reliability under low light is an empirical assumption deferred to Appendix Sec. IV, which is a validation gap rather than a circular step. The gating scalars lambda_s and lambda_g are trained on one dataset and then the model is evaluated on held-out zero-shot benchmarks (iSAID-dark high-pixel, UCM-RSLL, DarkRS, U3D, DOTA) using a model trained solely on iSAID-dark or LOLv1, so the headline improvements are not forced by fitting the test distribution. No load-bearing self-citation or imported uniqueness theorem appears; DINOv3 and Depth Anything 3 are external, non-author models. The 'Attention Drift' terminology is a post-hoc framing of known attention failure modes, but the paper does not use that framing to replace empirical evidence. Overall, no step reduces, by construction or by self-citation, to its own input.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The method's central effect depends on the standard degradation model, the robustness of frozen foundation model priors in extreme darkness, and the heuristic decomposition of attention drift. The five learned scalar gates and scales, plus the unspecified window size, are fitted to the training benchmarks. No new physical entities are proposed; attention drift is a named conceptual failure mode without independent measurement.

free parameters (6)

lambda_s (semantic gating scalar) = learned during training on iSAID-dark or LOLv1
Non-negative learnable global gate in Eq. (10) that weighs the semantic homogeneity bias; fitted to the target datasets.
lambda_g (geometric gating scalar) = learned during training on iSAID-dark or LOLv1
Non-negative learnable global gate in Eq. (10) that weighs the geometric heterogeneity penalty; fitted to the target datasets.
alpha (geometric penalty scale) = learned positive scale factor, value not reported
Scale factor in Eq. (9), B_geo(i,j) = -alpha |g_i - g_j|; controls the strength of the boundary penalty.
tau (semantic temperature) = learned temperature, value not reported
Temperature in Eq. (8), B_sem = Norm(S)Norm(S)^T / tau; controls sharpness of the semantic similarity distribution.
gamma (residual scale) = learned residual scale, value not reported
Learnable scale in Eq. (11), Z' = Z + gamma W^O Softmax(L)V; controls the magnitude of the attention update.
local window size (token count N per window) = not specified in main text
Hand-chosen architecture hyperparameter needed to reproduce the attention computation but omitted from the main text; details are likely in the unavailable appendix.

assumptions (4)

domain assumption Degradation model I(x) = J(x) ⊙ L(x) + N(x) with separable illumination and noise (Eq. 1).
Standard low-light imaging assumption used to motivate attention drift; the noise and illumination distributions are not specified, so Eq. (2) is heuristic.
domain assumption DINOv3 semantic features and Depth Anything 3 depth estimates computed from the corrupted input remain accurate enough to serve as hard attention constraints.
Load-bearing premise for H2CAM; validation is deferred to Appendix Sec. IV, which is not included in the reviewed text.
ad hoc to paper The degraded attention affinity can be approximated by Softmax(((q_J⊙l+n)(k_J⊙l+n)^T)/sqrt(d)) and split into Type I and Type II error terms.
Presented as a mathematical approximation in Eq. (2), but no derivation from a defined stochastic model is supplied; it is a motivating heuristic.
domain assumption The additive biases lambda_s B_sem and lambda_g B_geo act as hard physical constraints rather than soft regularizers because they enter before softmax.
The paper assumes deterministic bounds on aggregation; in practice the learned scalars can be small, and the constraints are trainable rather than fixed physical laws.

invented entities (1)

Attention Drift (with Type I Cross-Boundary Confusion and Type II Coplanar Ambiguity)
purpose: Names and organizes the failure modes that HALO is designed to solve; used to justify the dual-prior attention biases.
The phenomenon is plausible and illustrated with qualitative examples, but the paper provides no independent falsifiable measurement of attention drift separate from the proposed method's benchmark results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Overcoming Attention Drift: Homogeneity-Heterogeneity Guided Feature Aggregation for Low-Light Remote Sensing Image Enhancement." pith.science (2026). https://pith.science/paper/4DVCCTXK

@misc{pith2026260805843,
  author       = {Pith},
  title        = {Pith review of: Overcoming Attention Drift: Homogeneity-Heterogeneity Guided Feature Aggregation for Low-Light Remote Sensing Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4DVCCTXK}},
  note         = {Machine review of arXiv:2608.05843}
}

read the original abstract

Restoring high-fidelity remote sensing imagery from extreme low-light degradation is indispensable for reliable Earth observation and downstream machine vision. However, under severe noise and illumination corruption, existing methods suffer from attention drift, erroneously aggregating features across distinct physical boundaries and causing severe structural blurring and color distortion. To address this, we propose HALO, a dual-prior-driven enhancement framework that formulates enhancement as a guided feature aggregation problem driven by foundation model priors. Specifically, an illumination-invariant semantic prior provides regional homogeneity as a positive bias for content-consistent aggregation, while a pseudo-3D topological prior provides boundary heterogeneity as a negative penalty to strictly prevent cross-boundary confusion. To cooperatively incorporate these two priors, we propose a Homogeneity-Heterogeneity Cooperative Attention Module (H2CAM) to resolve feature conflicts during cross-modal prior fusion. Extensive experiments demonstrate that HALO achieves state-of-the-art performance across 8 challenging synthetic and real-world remote sensing benchmarks, significantly improving physical boundary sharpness and color fidelity while maximizing the preservation of discriminative features for downstream Earth observation tasks.

Figures

Figures reproduced from arXiv: 2608.05843 by the authors.

**Figure 2.** Performance illustration. (a) Low-light degradation causes feature entanglement. (b) and (c) Geometric and semantic priors provide reliable structural [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Overview of the HALO framework. (a) A cascaded pipeline restores features using dual-stream encoding and decoding. (b) Semantic and geometric [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Visual comparisons of the enhanced results by different methods on LOLv1 [ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Visual comparison of enhancement methods on U3D [ [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Visual comparison on the LSRW-Nikon dataset [ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Visual comparisons of object detection on the low-light DOTA-v1.0 dataset [ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 36 canonical work pages

[1]

Remote sensing image scene classifi- cation: Benchmark and state of the art,

G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classifi- cation: Benchmark and state of the art,”Proceedings of the IEEE, vol. 105, no. 10, pp. 1865–1883, 2017

2017
[2]

Deep learning in remote sensing: A comprehensive review and list of resources,

X. X. Zhu, D. Tuia, L. Mou, G.-S. Xia, L. Zhang, F. Xu, and F. Fraundorfer, “Deep learning in remote sensing: A comprehensive review and list of resources,”IEEE Geoscience and Remote Sensing Magazine, vol. 5, no. 4, pp. 8–36, 2017

2017
[3]

Dota: A large-scale dataset for object detection in aerial images,

G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “Dota: A large-scale dataset for object detection in aerial images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 3974–3983

work page 2018
[4]

Atmospheric scattering model and non-uniform illumination compensation for low-light remote sensing image enhancement,

X. Zhao, L. Huang, M. Li, C. Han, and T. Nie, “Atmospheric scattering model and non-uniform illumination compensation for low-light remote sensing image enhancement,”Remote Sensing, vol. 17, no. 12, p. 2069, 2025

work page 2025
[5]

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,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 3, 2023, pp. 2654–2662

2023
[6]

Deep retinex decomposition for low-light enhancement,

W. Chen, W. Wang, W. Yang, and J. Liu, “Deep retinex decomposition for low-light enhancement,” inProceedings of the British Machine Vision Conference, 2018

work page 2018
[7]

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,” inProceedings of the ACM International Conference on Multimedia, 2023, pp. 7459–7469

work page 2023
[8]

Spatial–frequency dual-domain feature fusion network for low-light remote sensing image enhancement,

Z. Yao, G. Fan, J. Fan, M. Gan, and C. L. P. Chen, “Spatial–frequency dual-domain feature fusion network for low-light remote sensing image enhancement,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–16, 2024

work page 2024

Show all 50 references

[9]

Low-light image and video enhancement using deep learning: A sur- vey,

C. Li, C. Guo, L. Han, J. Jiang, M.-M. Cheng, J. Gu, and C. C. Loy, “Low-light image and video enhancement using deep learning: A sur- vey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 12, pp. 9396–9416, 2022. 10

2022
[10]

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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5637–5646

2022
[11]

Uformer: A general u-shaped transformer for image restoration,

Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li, “Uformer: A general u-shaped transformer for image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 17 683–17 693

2022
[12]

Hvi: A new color space for low-light image enhancement,

Q. Yan, Y . Feng, C. Zhang, G. Pang, K. Shi, P. Wu, W. Dong, J. Sun, and Y . Zhang, “Hvi: A new color space for low-light image enhancement,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 5678–5687

2025
[13]

Structure- guided diffusion transformer for low-light image enhancement,

X. Yin, Z. Yu, L. Jiang, X. Gao, X. Sun, Z. Liu, and X. Yang, “Structure- guided diffusion transformer for low-light image enhancement,”IEEE Transactions on Multimedia, vol. 27, pp. 9505–9515, 2025

2025
[14]

Saigformer: A spatially-adaptive illumination-guided network for low-light image enhancement,

H. Li, F. Zhou, X. Sun, Y . Hua, J. Han, and L.-J. Zhang, “Saigformer: A spatially-adaptive illumination-guided network for low-light image enhancement,”arXiv preprint arXiv:2507.15520, 2025

2025 arXiv
[15]

Pixie: Prompted pixel-space low-light image enhancement,

R. Lin, G. Huang, D. Bull, and N. Anantrasirichai, “Pixie: Prompted pixel-space low-light image enhancement,”arXiv preprint arXiv:2605.23531, 2026

2026 arXiv
[16]

Sim ´eoni, H

O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, F. Massa, D. Haz- iza, L. Wehrstedt, J. Wang, T. Darcet, T. Moutakanni, L. Sentana, C. Roberts, A. Vedaldi, J. Tolan, J. Brandt, C. Couprie, J. Mairal,...

2025 arXiv
[17]

Learning semantic-aware knowledge guidance for low-light image en- hancement,

Y . Wu, C. Pan, G. Wang, Y . Yang, J. Wei, C. Li, and H. T. Shen, “Learning semantic-aware knowledge guidance for low-light image en- hancement,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 1662–1671

2023
[18]

Semantic-guided zero-shot learning for low- light image/video enhancement,

S. Zheng and G. Gupta, “Semantic-guided zero-shot learning for low- light image/video enhancement,” inProceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision Workshops, 2022, pp. 581–590

2022
[19]

Depth anything 3: Recovering the visual space from any views,

H. Lin, S. Chen, J. H. Liew, D. Y . Chen, Z. Li, Y . Zhao, S. Peng, H. Guo, X. Zhou, G. Shi, J. Feng, and B. Kang, “Depth anything 3: Recovering the visual space from any views,” inProceedings of the International Conference on Learning Representations, 2026

2026
[20]

Depth-aware blind image decomposition for real-world adverse weather recovery,

C. Wang, Z. Zheng, R. Quan, and Y . Yang, “Depth-aware blind image decomposition for real-world adverse weather recovery,” inProceedings of the European Conference on Computer Vision, 2024, pp. 379–397

2024
[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

2020
[22]

QWR-Dec-Net: A quaternion-wavelet retinex framework for low-light image enhancement with applications to remote sensing,

V . Frants, S. Agaian, K. Panetta, and A. Grigoryan, “QWR-Dec-Net: A quaternion-wavelet retinex framework for low-light image enhancement with applications to remote sensing,”Information, vol. 17, no. 1, p. 89, 2026

2026
[23]

Boosting diffusion networks with deep external context-aware encoders for low-light image enhancement,

P. Tang, Y . Wang, and A. Men, “Boosting diffusion networks with deep external context-aware encoders for low-light image enhancement,” Sensors, vol. 25, no. 23, p. 7232, 2025

2025
[24]

Spjfnet: Self-mining prior-guided joint frequency enhancement for ultra-efficient dark im- age restoration,

T. Zhang, P. Liu, Z. Zhang, and Q. Zhou, “Spjfnet: Self-mining prior-guided joint frequency enhancement for ultra-efficient dark im- age restoration,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 15, 2026, pp. 12 798–12 806

2026
[25]

Foundation models defining a new era in vision: A survey and outlook,

M. Awais, M. Naseer, S. Khan, R. M. Anwer, H. Cholakkal, M. Shah, M.-H. Yang, and F. S. Khan, “Foundation models defining a new era in vision: A survey and outlook,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 4, pp. 2245–2264, 2025

2025
[26]

Edge- connect: Structure guided image inpainting using edge prediction,

K. Nazeri, E. Ng, T. Joseph, F. Z. Qureshi, and M. Ebrahimi, “Edge- connect: Structure guided image inpainting using edge prediction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019, pp. 3265–3274

2019
[27]

Training-free large model priors for multiple- in-one image restoration,

X. He, L. Li, Y . Wang, H. Zheng, K. Cao, K. Yan, R. Li, C. Xie, J. Zhang, and M. Zhou, “Training-free large model priors for multiple- in-one image restoration,”arXiv preprint arXiv:2407.13181, 2024

2024 arXiv
[28]

DSPFusion: Image fusion via degradation and semantic dual-prior guidance,

L. Tang, C. Li, Y . Wang, G. Wang, Y . Yuan, and J. Ma, “DSPFusion: Image fusion via degradation and semantic dual-prior guidance,”IEEE Transactions on Image Processing, vol. 35, pp. 6331–6345, 2026

2026
[29]

TPGDiff: Hierarchical triple-prior guided diffusion for image restoration,

Y . Tu, Q. Yan, A. Niu, and J. Tang, “TPGDiff: Hierarchical triple-prior guided diffusion for image restoration,”arXiv preprint arXiv:2601.20306, 2026

2026 arXiv
[30]

Multiprior learning via neural architecture search for blind face restoration,

Y . Yu, P. Zhang, K. Zhang, W. Luo, and C. Li, “Multiprior learning via neural architecture search for blind face restoration,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 2, pp. 3057– 3070, 2025

2025
[31]

Lersgan: A gan-based model for low-light remote sensing image enhancement,

T. Li, T. Guo, and D. Xiang, “Lersgan: A gan-based model for low-light remote sensing image enhancement,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 18, pp. 26 489– 26 504, 2025

2025
[32]

Deepspg: Exploring deep semantic prior guidance for low-light image enhancement with multi- modal learning,

J. Lu, H. Zhao, H. Zhai, X. Yang, and S. Han, “Deepspg: Exploring deep semantic prior guidance for low-light image enhancement with multi- modal learning,” inProceedings of the ACM International Conference on Multimedia Retrieval, 2025, pp. 935–943

2025
[33]

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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5891–5900

2022
[34]

Empowering low-light image enhancer through customized learnable priors,

N. Zheng, M. Zhou, Y . Dong, X. Rui, J. Huang, C. Li, and F. Zhao, “Empowering low-light image enhancer through customized learnable priors,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 12 525–12 535

2023
[35]

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

2023
[36]

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,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 12 872–12 881

2023
[37]

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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 22 252–22 261

2023
[38]

Low-light image enhancement via generative perceptual priors,

H. Zhou, W. Dong, X. Liu, Y . Zhang, G. Zhai, and J. Chen, “Low-light image enhancement via generative perceptual priors,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 10, 2025, pp. 10 752–10 760

2025
[39]

Bayesian neural networks for one-to-many mapping in image enhance- ment,

G. Huang, Q. Yang, R. Lin, Z. Qi, D. Bull, and N. Anantrasirichai, “Bayesian neural networks for one-to-many mapping in image enhance- ment,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 7, 2026, pp. 5004–5012

2026
[40]

Bag-of-visual-words and spatial extensions for land-use classification,

Y . Yang and S. Newsam, “Bag-of-visual-words and spatial extensions for land-use classification,” inProceedings of the ACM SIGSPATIAL Inter- national Conference on Advances in Geographic Information Systems, 2010, pp. 270–279

2010
[41]

R2rnet: Low-light image enhancement via real-low to real-normal network,

J. Hai, Z. Xuan, R. Yang, Y . Hao, F. Zou, F. Lin, and S. Han, “R2rnet: Low-light image enhancement via real-low to real-normal network,” Journal of Visual Communication and Image Representation, vol. 90, p. 103712, 2023

2023
[42]

Unsupervised ultra-high-resolution uav low-light image enhancement: A benchmark, metric and framework,

W. Lu, L. Zhu, and S.-B. Chen, “Unsupervised ultra-high-resolution uav low-light image enhancement: A benchmark, metric and framework,” arXiv preprint arXiv:2509.01373, 2025

2025 arXiv
[43]

Image quality assessment: from error visibility to structural similarity,

Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

2004
[44]

The unreasonable effectiveness of deep features as a perceptual metric,

R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 586–595

2018
[45]

No-reference image quality assessment in the spatial domain,

A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference image quality assessment in the spatial domain,”IEEE Transactions on Image Processing, vol. 21, no. 12, pp. 4695–4708, 2012

2012
[46]

Making a completely blind image quality analyzer,

A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a completely blind image quality analyzer,”IEEE Signal Processing Letters, vol. 20, no. 3, pp. 209–212, 2013

2013
[47]

Blind image quality evaluation using perception based features,

N. Venkatanath, D. Praneeth, B. M. Chandrasekhar, S. S. Channappayya, and S. S. Medasani, “Blind image quality evaluation using perception based features,” inProceedings of the 21st National Conference on Communications, 2015, pp. 1–6

2015
[48]

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, 2017

2017
[49]

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,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 8, pp. 4225–4238, 2022

2022
[50]

Poly kernel inception network for remote sensing detection,

X. Cai, Q. Lai, Y . Wang, W. Wang, Z. Sun, and Y . Yao, “Poly kernel inception network for remote sensing detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 27 706–27 716

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Remote sensing image scene classifi- cation: Benchmark and state of the art,

G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classifi- cation: Benchmark and state of the art,”Proceedings of the IEEE, vol. 105, no. 10, pp. 1865–1883, 2017

2017

[2] [2]

Deep learning in remote sensing: A comprehensive review and list of resources,

X. X. Zhu, D. Tuia, L. Mou, G.-S. Xia, L. Zhang, F. Xu, and F. Fraundorfer, “Deep learning in remote sensing: A comprehensive review and list of resources,”IEEE Geoscience and Remote Sensing Magazine, vol. 5, no. 4, pp. 8–36, 2017

2017

[3] [3]

Dota: A large-scale dataset for object detection in aerial images,

G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “Dota: A large-scale dataset for object detection in aerial images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 3974–3983

work page 2018

[4] [4]

Atmospheric scattering model and non-uniform illumination compensation for low-light remote sensing image enhancement,

X. Zhao, L. Huang, M. Li, C. Han, and T. Nie, “Atmospheric scattering model and non-uniform illumination compensation for low-light remote sensing image enhancement,”Remote Sensing, vol. 17, no. 12, p. 2069, 2025

work page 2025

[5] [5]

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,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 3, 2023, pp. 2654–2662

2023

[6] [6]

Deep retinex decomposition for low-light enhancement,

W. Chen, W. Wang, W. Yang, and J. Liu, “Deep retinex decomposition for low-light enhancement,” inProceedings of the British Machine Vision Conference, 2018

work page 2018

[7] [7]

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,” inProceedings of the ACM International Conference on Multimedia, 2023, pp. 7459–7469

work page 2023

[8] [8]

Spatial–frequency dual-domain feature fusion network for low-light remote sensing image enhancement,

Z. Yao, G. Fan, J. Fan, M. Gan, and C. L. P. Chen, “Spatial–frequency dual-domain feature fusion network for low-light remote sensing image enhancement,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–16, 2024

work page 2024