acceptodds
Withdrawn from ICLR 2027

DPFR-LLIE: Dual-Prior Frequency Routing for UAV Low-Light Image Enhancement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

UAV images captured from high altitudes and long viewing distances often suffer from uneven illumination, low signal-to-noise ratios, color distortion, and platform motion. These coupled degradations suppress object boundaries and textures, leaving the degraded observation with insufficient evidence to reliably recover the structural details of small and distant targets. Existing low-light enhancement methods infer restoration cues mainly from this incomplete observation, therefore weak structural information can be mistaken for noise or blur and smoothed away. To obtain structural guidance that generalizes across low-light degradations, we analyze frozen DINOv3 feature representations learned from large-scale visual pretraining. We find that they retain cross-exposure structural correspondence under severe degradation, with low-, mid-, and high-frequency components showing their strongest correspondence at different feature depths. Based on this feature-depth–frequency dependency, we propose Dual-Prior Frequency Routing for Low-Light Image Enhancement (DPFR-LLIE). We first design a frequency-specific feature-depth aggregation module to learn band-specific mixtures of frozen DINOv3 features. Then we design a bounded input-conditioned routing module to allocate and inject the resulting structural guidance with controlled magnitude. Finally, we propose a photometric conditioning module to modulate encoder features with image-derived luminance and chrominance cues. Across the UAV scenes, DPFR-LLIE achieves the best quantitative scores and the strongest qualitative results among the evaluated methods. Specifically, compared with the current SOTA method StarIR, DPFR-LLIE improves PSNR/SSIM by 0.71 dB/0.014 on Synthetic VisDrone Low-Light and by 1.42 dB/0.022 on LOL-Blur. With the same frozen detector, the enhanced UAV images further improve AP and by 0.27 and 0.31 points. The code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
low-light image enhancement, UAV vision, DINOv3, frequency-aware learning, dynamic routing, image restoration

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_Kh7MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DPFR-LLIE for UAV low-light image enhancement. It combines image-derived photometric priors with structural guidance from a frozen DINOv3 encoder. Features from different encoder depths are aggregated separately for low-frequency, mid-frequency, and high-frequency bands, followed by input-dependent routing into the restoration network. The method is evaluated on Synthetic VisDrone, LOL-Blur, and real nighttime UAV images from DarkRS, with additional evaluation using a fixed object detector.

Critical strengths
  1. The method and implementation are reasonably complete. The paper describes feature extraction, frequency decomposition, feature-depth aggregation, routing supervision, and residual injection, with corresponding implementations in the supplementary code. The ablation protocols also clarify the training settings and retained auxiliary objectives.
  2. The complete system shows positive results in both restoration and downstream detection. The reported PSNR improvements over StarIR are 0.71 dB on Synthetic VisDrone and 1.42 dB on LOL-Blur.
Critical weaknesses
  1. The main SSIM results are internally inconsistent. Weighting the subgroup SSIM values in Table 8 by the validation counts in Table 6 gives approximately 0.8645, rather than the reported overall value of 0.867. This discrepancy exceeds the possible rounding error, and the supplied evaluation code averages per-image SSIM values.
  2. The incremental contribution of frequency aggregation and dynamic routing remains limited. The complete DINO branch adds 0.12 dB over the photometric-only model, while dynamic routing adds 0.08 and 0.06 dB over uniform and static allocation. Gains over shared feature-depth fusion are similarly small, leaving limited evidence of a substantial advance over simpler alternatives.
Questions for authors
  1. Could you reconcile the reported overall SSIM of 0.867 with the subgroup-weighted value of approximately 0.8645? Please verify the evaluation sets, checkpoints, and metric settings using the per-image records, and clarify whether any ablation results require correction.
  2. How does the benefit of dynamic routing over static allocation vary across degradation levels? Could you provide a breakdown of the existing per-image results?
Reviewer_kRakOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DPFR-LLIE for low-light enhancement of UAV images, targeting photometric correction and the preservation of weak structures in small and distant objects. The authors analyze frozen DINOv3 features and observe that correspondence between degraded and reference images varies with feature depth and frequency band. Motivated by this observation, the method learns separate feature-depth mixtures for low-, mid-, and high-frequency guidance, followed by bounded input-conditioned routing with Equal-RMS control of the injected residual. A complementary photometric conditioning module modulates encoder features using image-derived luminance and chrominance cues. DPFR-LLIE improves PSNR over StarIR by 0.71 dB on Synthetic VisDrone and 1.42 dB on LOL-Blur, and yields modest improvements in object detection with a fixed detector. Transfer experiments on real nighttime UAV images show competitive no-reference quality, although the advantage is less clear than on the paired benchmarks.

Critical strengths
  1. The complete system achieves consistent improvements on two paired datasets and includes downstream-task validation. Relative to StarIR, DPFR-LLIE improves PSNR and SSIM by 0.71 dB and 0.014 on Synthetic VisDrone, and by 1.42 dB and 0.022 on LOL-Blur. Improvements hold across all four degradation levels on Synthetic VisDrone. With a fixed Faster R-CNN detector and only the enhanced inputs changed, AP and small-object AP increase by 0.27 and 0.31 percentage points, respectively. Although the detection gains are modest, these results support some downstream utility beyond improvements in image reconstruction metrics.

  2. The ablations carefully control several important confounding factors. The paper separately evaluates photometric conditioning, DINO guidance, and their combination, and provides matched comparisons of uniform, static, and dynamic routing. In the routing controls, the auxiliary predictor and its supervision remain active, while Equal-RMS matching controls the aggregate magnitude of the injected residual. These choices reduce confounding from auxiliary supervision and changes in feature magnitude. The authors also report paired per-image performance differences and explicitly acknowledge when a control changes both the representation and its supervision. This improves the interpretability of the experiments and helps distinguish complete-system gains from individual component contributions.

Critical weaknesses
  1. The additional gains from the central frequency-routing and feature-depth aggregation designs are small, and their reliability and significance remain insufficiently established. Table 2 shows that photometric conditioning alone raises PSNR from 26.10 to 26.69 dB, while adding the complete DINO branch and its associated supervision further increases it to 26.81 dB. Dynamic routing improves PSNR over static routing by only approximately 0.06 dB in Table 3, while SSIM decreases slightly from 0.868 to 0.867. Figure 4 further shows that dynamic routing outperforms static routing on only 49.3% of validation images. Separate feature-depth mixtures for different frequency bands likewise improve PSNR over shared fusion by approximately 0.06 dB. These results indicate limited average gains, but the paper does not report variability across independent training runs or confidence intervals for these differences, and the key ablations are restricted to Synthetic VisDrone. Consequently, it remains unclear whether the small advantages are stable, generalize to another dataset, and substantiate the importance of the proposed frequency-guidance design. Paired per-image comparisons are informative but do not establish robustness to training randomness.

  2. Evidence of generalization to independent UAV scenes and real nighttime degradations remains limited. Appendix C states that the 2,059 validation pairs correspond to 528 distinct reference images, with multiple degradation variants sharing a reference image. The authors verify that reference filenames do not overlap between training and validation, but explicitly note that this does not establish disjoint source videos or scenes. This leaves uncertainty about whether the principal UAV experiment adequately tests generalization to independent scenes; it does not establish that leakage has occurred. On the real nighttime UAV dataset DarkRS, quantitative evaluation relies on NIQE. DPFR-LLIE scores 4.129, slightly better than StarIR's 4.184 but worse than Multinex's 3.959, without detection results or a direct assessment of structural fidelity on that dataset. The evidence therefore supports reconstruction improvements under synthetic degradation more strongly than the recovery of weak small-object structures in real nighttime UAV imagery, which is a central motivation of the paper.

Questions for authors
  1. Reliability of the core component gains. Dynamic routing over static routing and separate feature-depth fusion over shared fusion each improve PSNR by approximately 0.06 dB. Are results from independent training runs or confidence intervals available to establish that these differences exceed training variability? Do the same advantages hold on LOL-Blur? This evidence would affect the assessment of the central modules' contributions.

  2. Scene independence of the dataset split. Appendix C confirms that reference filenames do not overlap between training and validation. Can the authors clarify whether the underlying videos or scenes are also disjoint, and whether source-sequence information was used when constructing the split? If this has not been verified, are results available under a video- or scene-disjoint split to determine whether the reported advantages persist?

  3. Structural preservation on real nighttime images. Beyond NIQE and qualitative examples on DarkRS, is there evidence from object detection, human assessment of structural fidelity, or another direct evaluation that supports the recovery of weak small-object structures in real nighttime images? This would help establish whether the synthetic-benchmark improvements extend to the intended application setting.

Reviewer_ro62OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DPFR-LLIE, a dual-prior framework for UAV low-light image enhancement. The method exploits frozen DINOv3 features to provide structural guidance and introduces frequency-specific feature-depth aggregation to combine representations from different layers. A bounded dynamic routing module adaptively integrates frequency-specific features, while a photometric conditioning module provides luminance and chrominance guidance. Experiments on Synthetic VisDrone Low-Light, LOL-Blur, and DarkRS demonstrate improvements in image enhancement quality, with additional downstream object detection evaluations.

Critical strengths
  1. Motivated integration of structural and photometric priors. The paper provides an interesting analysis of frequency-dependent representations across DINOv3 feature depths. Combining pretrained structural guidance with image-derived photometric information is a reasonable approach to addressing illumination degradation and structural detail loss in UAV imagery.

  2. Competitive enhancement performance with downstream evaluation. The proposed method achieves improvements over existing methods on Synthetic VisDrone and LOL-Blur. The authors also investigate several architectural components through ablation experiments and evaluate the influence of enhanced images on downstream object detection.

  3. Could the authors provide more detailed comparisons with recent DINO-based restoration approaches, particularly DINOLight and OmniLight, to clarify the methodological novelty?

  4. Given the relatively small improvements from dynamic routing and DINO guidance, can the authors provide additional evaluations demonstrating their consistent benefits across different degradation levels and datasets?

  5. Could the authors report the complete computational cost, including DINOv3 feature extraction, inference latency, and peak GPU memory usage?

  6. Can the authors provide stronger quantitative evidence on real-world nighttime UAV imagery to demonstrate generalization beyond synthetic degradations?

  7. Was the main framework diagram in Figure 2 created or substantially modified using generative AI?

Critical weaknesses
  1. Limited methodological novelty. The proposed framework primarily combines pretrained DINOv3 features, frequency decomposition, feature aggregation, and dynamic routing. Similar ideas have been explored in existing restoration methods, such as DINOLight and OmniLight. The fundamental methodological distinction is not sufficiently established.
  2. Marginal contributions of the core modules. According to Tables 2–4, the proposed DINO guidance and dynamic routing provide relatively small improvements over simpler alternatives. In particular, the full model improves PSNR by only 0.12 dB over photometric conditioning alone, while dynamic routing offers just 0.06 dB over static routing. These modest gains weaken the justification for the additional architectural complexity.
  3. Insufficient validation on real-world UAV images. Most quantitative evidence comes from synthetically degraded images or LOL-Blur. The real-world DarkRS evaluation relies only on NIQE, where the proposed method does not achieve the best result. Stronger evaluations on real nighttime UAV images are necessary to substantiate its practical effectiveness.
  4. Missing computational complexity and efficiency analysis. The framework incorporates a frozen DINOv3 encoder and multiple frequency-processing operations, but does not report comprehensive parameter counts, FLOPs, inference latency, or GPU memory consumption. The additional computational burden is particularly important for practical UAV deployment.
  5. Concerns regarding the potentially AI-generated main framework figure. The main framework diagram in Figure 2 appears to be AI-generated.
  6. Incomplete comparison with closely related foundation-model methods. Although DINOLight, CANDLE, and OmniLight are discussed in the related work, they are not included in the quantitative comparisons. Without evaluating these closely related approaches, it is difficult to establish the advantages of the proposed DINO-based frequency routing strategy.
Questions for authors
  1. Could the authors provide more detailed comparisons with recent DINO-based restoration approaches, particularly DINOLight and OmniLight, to clarify the methodological novelty?

  2. Given the relatively small improvements from dynamic routing and DINO guidance, can the authors provide additional evaluations demonstrating their consistent benefits across different degradation levels and datasets?

  3. Could the authors report the complete computational cost, including DINOv3 feature extraction, inference latency, and peak GPU memory usage?

  4. Can the authors provide stronger quantitative evidence on real-world nighttime UAV imagery to demonstrate generalization beyond synthetic degradations?

  5. Was the main framework diagram in Figure 2 created or substantially modified using generative AI?

Reviewer_448bOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DPFR-LLIE for UAV low-light image enhancement (LLIE). The proposed method consists of a three-stage U-Net backbone, a photometric conditioning module that separately extracts luminance and chrominance features before concatenating them, and a routed structural guidance module based on a frozen DINOv3 (ViT-S/16) encoder. Despite targeting UAV LLIE, the evaluation is conducted primarily on a synthetic ground-level dataset (LOL-Blur) and a self-constructed synthetically degraded UAV dataset (VisDrone). The authors also conduct a zero-shot evaluation on DarkRS using models trained on LOL-Blur rather than those trained on the UAV-oriented VisDrone dataset. This evaluation protocol is somewhat questionable, and the rationale behind this choice is not sufficiently explained. Although the reported results appear promising, the experimental comparison does not include several recent state-of-the-art LLIE methods, such as HVI-CIDNet (CVPR'25), DarkIR (CVPR'25), GM-MoE (ICCV'25), MR. Illuminate (CVPR'26), and Bi-Bridge (CVPR'26). The lack of comparisons with these recent methods makes it difficult to assess the actual performance and competitiveness of DPFR-LLIE.

Critical strengths

High performance on LOL-Blur dataset. Although not in their field (UAV Data).

Critical weaknesses
  1. Insufficient quantitative and qualitative evaluation on real-world UAV data.

(1) The real-world dataset DarkRS used in this paper contains only 86 images, which is insufficient to provide convincing evidence of the method's effectiveness and robustness in real-world UAV scenarios. Furthermore, the visual comparisons in Fig. 7 suggest that the proposed method may introduce more false-color artifacts than RetinexFormer, raising concerns about its color fidelity.

(2) The authors should conduct zero-shot evaluation on DarkRS using the pretrained model trained on their synthetic VisDrone dataset. The current evaluation uses a model trained on LOL-Blur instead, leaving the transferability of the proposed UAV-oriented training dataset unexplored. This raises questions about the actual utility and contribution of the synthetic VisDrone dataset to the proposed method.

  1. Limited methodological novelty.

The proposed method appears to primarily integrate existing techniques, including multi-layer DINO feature aggregation, Fourier-based frequency decomposition, dynamic routing, and affine feature modulation. However, the paper does not clearly identify what fundamentally new methodological insight or technical contribution is introduced beyond this combination of existing components. The authors should better clarify the novelty of the proposed framework and distinguish it from prior work.

Questions for authors

I do not see sufficient justification for treating UAV LLIE as a distinct task from general LLIE. The paper does not adequately identify the unique challenges of low-light image enhancement in UAV scenarios or explain why a dedicated UAV-oriented approach is necessary. The authors should either reconsider the task positioning and resubmit the work as a general LLIE paper, or provide a carefully collected and representative UAV LLIE dataset, together with a thorough analysis demonstrating how the UAV imaging domain differs from ground-level imaging and why these differences necessitate dedicated solutions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.