PaperlayerView exact revision on arXiv

Source-linked AI summary

Physics-Guided Spectral Distillation for Underwater Image Enhancement on Resource-Constrained Devices

Yifan Chen, Kai He, Ye Zheng, Jijun Lu, Zhe Sun, Tao Chen

arXiv:2609.34795v1cs.CV

TL;DR

Underwater image enhancement must address coupled color, contrast, and detail degradation, while compact models are needed for resource-constrained applications. PSD distills band-specific, degradation-aware and reliability-gated guidance into compact students. Across three backbone–dataset settings, it reports large parameter reductions with small average PSNR and SSIM drops, alongside downstream detection and ROV deployment results.

  • Problem

    Underwater images suffer coupled color, contrast, and detail degradation, while existing distillation objectives may not separately address frequency-specific restoration, spatially varying attenuation, and unreliable teacher predictions.

    Source: §I · pp. 1–2

  • Method

    PSD distills compact UIE students using multilevel wavelet bands, transmission-derived degradation weights, ground-truth-guided reliability masks, and image-space compensation.

    Source: §I · p. 2

  • Results

    Across three backbone–dataset settings, compact students reduce teacher parameter counts by 95.45%–99.06% with average drops of only 0.501 dB in PSNR and 0.0059 in SSIM.

    Source: §I · p. 2

  • Takeaways & Limitations

    UDD detection experiments and self-developed ROV deployment show downstream benefits and practical applicability in real-world underwater scenarios.

    Source: §I · p. 2, §V · p. 9

Abstract

from arXiv · show

Underwater image enhancement is crucial for improving visual perception in marine applications. Existing underwater image enhancement studies mainly focus on enhancement quality and visual fidelity, while rarely considering real-time deployment capability, which is essential for resource-constrained underwater robots. To this end, we introduce a physics-guided spectral distillation (PSD) method, which reduces model capacity for real-time applications while maintaining the high performance of underwater image enhancement models. To decompose the outputs of teacher and student models, PSD adopts a multilevel Haar discrete wavelet transform. It transfers low-frequency color and illumination information as well as high-frequency structural details through band-specific objectives. Moreover, the distillation process of PSD is degradation-aware. We estimate degradation-aware weights through a physical head and combine them with ground-truth-guided reliability masks to selectively retain valuable teacher guidance. Experiments on the UIEB, LSUI, and EUVP datasets validate the effectiveness of the proposed method. Furthermore, we demonstrate the benefits of enhanced images for downstream perception tasks, including object detection. Deployment on a self-developed ROV further demonstrates its practical applicability in real-world underwater scenarios.

I. INTRODUCTION

Underwater image enhancement must correct coupled color, contrast, and detail degradation while meeting the limited computation, memory, and power budgets of underwater robots. PSD uses band-specific, degradation-aware distillation to improve compact students, with reported benchmark, detection, and deployment results.

I. INTRODUCTION

  • Underwater robots must process images within limited computation, memory, and power budgets while preserving quality at low inference cost.Underwater absorption and scattering distort color, reduce contrast, and blur details across real-world water and illumination conditions.Source: §I · p. 1
  • Traditional handcrafted or physical-model methods can leave residual color casts or overcorrect when scene statistics, water properties, or illumination differ from their assumptions.Learned models offer richer restoration mappings, but their compute and memory demands can limit embedded real-time processing; narrowing them can weaken color recovery and boundary preservation.Source: §I · p. 1
  • Compact UIE students need band-specific supervision that accounts for degradation severity and teacher reliability because image-space matching misses spectral distinctions and teacher predictions can be inaccurate regionally.Downsampling examples show progressive high-frequency detail loss.Source: §I · pp. 1–2, Figure 2 · p. 2
  • PSD decomposes outputs with a multilevel Haar transform, weights degraded regions and channels, gates spectral targets by teacher advantage, and adds an unmasked RGB loss.The teacher, physical head, and auxiliary computations are used during training but add no inference-time modules.Source: §I · p. 2
  • PSD reduces teacher parameter counts by 95.45%–99.06%, with average drops of only 0.501 dB in PSNR and 0.0059 in SSIM, while improving compact students at unchanged inference cost.The accuracy–efficiency comparison covers Reti-Diff, Restormer, and NAFNet students.Source: Figure 1 · p. 1, §I · p. 2
  • Detection gains over raw inputs reach 3.22 and 1.85 percentage points in mAP50 and mAP50:95, respectively, and students achieve 35.7–84.6 FPS at 256 × 256 resolution.The introduction reports downstream validation on UDD and practical deployment on a self-developed ROV.Source: §I · p. 2

II. RELATED WORK

Prior underwater enhancement work spans physical models, learned restoration, and lightweight architectures, while distillation research offers several ways to transfer knowledge. PSD's overview applies degradation-aware, reliability-gated spectral transfer to compact UIE models and removes its training auxiliaries before deployment.

A. Underwater Image Enhancement

  • UIE research spans physical imaging models, learned restoration, frequency and diffusion designs, and lightweight networks aimed at embedded platforms.Physical approaches depend on scene-appropriate statistics or optical parameters, while iterative or multi-branch designs can increase computation and memory.Source: A. Underwater Image Enhancement · p. 2

B. Knowledge Distillation

  • Distillation transfers outputs, features, attention, geometry, degradation, or spectral information, but existing selection rules do not encode channel-dependent underwater attenuation.PSD couples transmission-derived weighting with ground-truth-guided reliability masks and an unmasked RGB loss.Source: B. Knowledge Distillation · p. 2

A. Problem Formulation and Overview

  • PSD formulates training around frozen-teacher and compact-student outputs, estimates degradation, decomposes outputs into wavelet bands, and gates spectral transfer by the teacher’s advantage.Student training combines task, spectral, and image-space losses; only the compact student is deployed, with no PSD-specific inference modules.Source: A. Problem Formulation and Overview · p. 3, Figure 3 · p. 3

B. Underwater-Formation-Guided Degradation Weighting

A lightweight physical head is calibrated using a channel-dependent underwater image-formation model before student training. The training procedure freezes the calibrated head, uses its transmission estimate for PSD weights, and removes it after training.

B. Underwater-Formation-Guided Degradation Weighting

  • The channel-dependent image-formation model uses transmission and background light, which the lightweight physical head predicts from the degraded input.The model’s transmission is channel-wise; the background light is global.Source: 1) Physical-head calibration · p. 3
  • The channel-wise transmission estimate is bounded between 0.05 and 0.99 during physical-head calibration.Source: 1) Physical-head calibration · p. 3
  • Calibration reconstructs the degraded observation, then evaluates a physical calibration objective.The reference image is detached during calibration so only the physical head is updated.Source: 1) Physical-head calibration · p. 3
  • For standalone Restormer and NAFNet implementations, calibration adds a gradient-matching term weighted by 0.1.The auxiliary term stabilizes the physical estimator and does not supervise the student.Source: 1) Physical-head calibration · p. 3
  • After calibration, the physical head remains frozen while the student is trained, and both the teacher and physical head are removed afterward.The procedure computes transmission-derived weights and Haar bands during student training.Source: 1) Physical-head calibration · pp. 3–4
  • The training procedure uses paired degraded and reference images, a frozen teacher, a student, and the native task loss alongside the PSD loss.The appendix labels this procedure the complete PSD training algorithm.Source: 1) Physical-head calibration · p. 4

2) Transmission-derived degradation weight:

PSD turns calibrated transmission into channel-wise spatial weighting: low transmission marks stronger attenuation. The normalized weights are detached before use in distillation.

2) Transmission-derived degradation weight:

  • Low transmission indicates strong attenuation, so PSD converts calibrated transmission into a channel-wise spatial weight.Source: 2) Transmission-derived degradation… · p. 4
  • The weight map is normalized to fix its mean scale, uses γ = 2, and is detached before the distillation losses.Background light supports image-formation calibration, whereas transmission supplies PSD’s spatial and channel weighting.Source: 2) Transmission-derived degradation… · p. 4

C. Reliability-Aware Multilevel Spectral Distillation

PSD applies a two-level Haar decomposition to teacher, student, and reference outputs, separating coarse appearance from directional structure. Ground-truth-guided masks restrict transfer to coefficients where the teacher is closer to the reference than the student.

1) Two-level Haar representation:

  • The channel-preserving Haar transform is applied to restored outputs, with low-frequency bands representing coarse intensity and chromatic content and signed bands representing directional structures.The signed bands correspond to horizontal, vertical, and diagonal structure.Source: 1) Two-level Haar representation · p. 4
  • The decomposition starts with each output as its level-zero low-frequency band and recursively produces the next level’s low- and high-frequency bands.PSD uses K = 2 decomposition levels; odd spatial dimensions are cropped by one pixel before each transform.Source: 1) Two-level Haar representation · p. 4

2) Ground-truth-guided teacher-advantage masks:

  • PSD compares detached teacher and student errors against the paired reference to form low- and high-frequency reliability masks.The high-frequency comparison is channel-wise at each level.Source: 2) Ground-truth-guided… · p. 4
  • A coefficient contributes to transfer only when the teacher is closer to the reference than the current student, and the masks are recomputed as the student changes.Source: 2) Ground-truth-guided… · p. 4

3) Band-specific spectral objectives:

PSD weights low- and high-frequency transfer differently across bands. The low-frequency objective retains channel-wise weighting, while the high-frequency objective uses channel-averaged weights and downweights coarser levels.

3) Band-specific spectral objectives:

  • PSD downsamples the degradation weights to each subband’s resolution, retaining channel-wise weights for the final LL band and channel-averaged weights for high-frequency transfer.The LL term transfers coarse intensity and channel-dependent chromatic correction.Source: 3) Band-specific spectral objectives · p. 4
  • The high-frequency objective reduces coarser-level contributions with ρ = 0.5, and PSD sets αLL = 1 and αHF = 0.5.Source: 3) Band-specific spectral objectives · p. 4

D. Image-Space Compensation

Image-space compensation complements wavelet supervision with global RGB alignment, while reliability computation preserves the spectral coefficients’ native channel structure.

D. Image-Space Compensation

  • A low-weight image-space loss aligns student and teacher RGB outputs without a reliability mask.Reliability gating applies only to the LL and HF spectral losses.Source: D. Image-Space Compensation · p. 4
  • The reliability computation preserves the native channel structure of the spectral coefficients.Algorithm 2 summarizes this computation for the spectral objectives.Source: D. Image-Space Compensation · p. 5
  • The image-space term averages squared differences across RGB channels, pixels, and batch samples, with teacher outputs detached from gradient computation.It supplies global output alignment alongside selectively gated spectral supervision.Source: D. Image-Space Compensation · p. 5

IV. EXPERIMENTS AND APPLICATION VALIDATION

The experiments use paired-reference datasets and matched comparisons across three restoration backbones. Students retain each framework’s training strategy, while PSD-specific components are used only during training.

1) Datasets and evaluation metrics:

  • UIEB contributes 890 reference-paired images, LSUI 4,279 pairs, and merged EUVP data use an 8:2 training-validation split; PSNR and SSIM are measured against RGB references.All experiments use one fixed random seed without repeated-run confidence intervals.Source: 1) Datasets and evaluation metrics · p. 5

2) Compared distillation methods:

  • Teachers, task-only students, and four distillation methods share checkpoints, architectures, data splits, task losses, and training budgets within each backbone–dataset setting.The comparisons include CTKD, DCKD, FreeKD+, and PSD.Source: 2) Compared distillation methods · p. 5

3) Cross-backbone students and training protocol:

  • The compact students retain framework-specific training strategies, and PSD adds two-level spectral and image-space losses without feature adapters or inference-time components.The tested backbones are Reti-Diff, Restormer, and NAFNet.Source: 3) Cross-backbone students and… · p. 5

B. Model Compression and Computational Cost

Across three backbones and datasets, PSD narrows compact students’ teacher-performance gaps despite substantial model compression. Leave-one-component-out results show that all components contribute, with sensitivities varying by backbone.

B. Model Compression and Computational Cost

C. Comparison with Teachers and Distillation Methods

  • PSD’s mean PSNR/SSIM gaps were 1.350/0.0031 for Reti-Diff, 0.287/0.0108 for Restormer, and −0.133/0.0038 for NAFNet.PSD led all six NAFNet student comparisons; on LSUI it exceeded teacher PSNR by 0.44 dB but had 0.0082 lower SSIM.Source: C. Comparison with Teachers and… · p. 5
  • 0.501 dB PSNR and 0.0059 SSIM were the mean signed teacher–student gaps with PSD, down from 1.191 dB and 0.0156 across nine settings.PSD improved both metrics over task-only training in every setting and led 12 of 18 compact-student comparisons, with 95.45%–99.06% fewer parameters.Source: C. Comparison with Teachers and… · p. 5, Table II · p. 6

D. Ablation Study

  • Removing any single component reduced both metrics, supporting contributions from physical weighting, spectral transfer, reliability gating, and image-space compensation.The largest reported sensitivities were Restormer’s 0.40 dB PSNR drop without physical weighting and NAFNet’s 0.61 dB and 0.0066 SSIM drops without image-space compensation.Source: D. Ablation Study · p. 6
  • Removing HF transfer reduced SSIM for every backbone; Reti-Diff’s largest SSIM drop, 0.0026, followed removal of reliability gating.The results support complementary appearance, structural, and reliability constraints.Source: D. Ablation Study · p. 6

E. Downstream Object Detection

Enhanced image domains improve aggregate detection metrics over raw images, and the compact students also run at real-time rates in the reported protocol. ROV tests on sampled lake imagery show gains in feature matches and no-reference image-quality scores, with limits on what those measures establish.

E. Downstream Object Detection

  • A separate YOLOv9s detector was trained and tested in each raw or enhanced image domain under identical splits and a common protocol.The UDD evaluation used 2,227 images of sea cucumbers, sea urchins, and scallops.Source: E. Downstream Object Detection · p. 6
  • Figure 4 compares input, task-only student, teacher, CTKD, DCKD, FreeKD+, PSD, and ground-truth images across Reti-Diff, Restormer, and NAFNet.Figure 5 compares task-only and PSD students with the teacher and ground truth; red-box crops highlight local contrast, color, and structure.Source: Figure 4 · p. 7, Figure 5 · p. 7
  • Among PSD students, Restormer achieved the largest mAP50 gain (3.22 points), while NAFNet achieved the largest mAP50:95 gain (1.85 points).All six enhanced domains improved both aggregate metrics over Raw.Source: E. Downstream Object Detection · pp. 6–7, Table IV · p. 8
  • UDD’s uneven category gains are consistent with class imbalance; enhancement cannot replace missing supervision, and scallop AP does not improve consistently.Sea urchins dominate the dataset, while the other categories have fewer training examples and less stable AP estimates.Source: E. Downstream Object Detection · p. 7

F. Application on a Self-Developed ROV

  • The deployment used a self-developed ROV to collect images in Dushu Lake, Suzhou, China.The setup included an onboard GoPro and stereo camera.Source: F. Application on a Self-Developed ROV · p. 7
  • At 256×256, the students reached 35.7–84.6 FPS and 3.30–8.30× teacher speedups; all exceeded 30 FPS under this protocol.NAFNet was selected for field deployment because it had the highest throughput, despite having more parameters than Restormer.Source: F. Application on a Self-Developed ROV · p. 7
  • Across four frame pairs, SIFT matches increased from 2, 7, 12, and 26 to 8, 17, 36, and 52, raising the mean from 11.75 to 28.25 (2.40×).More correspondences suggest improved feature visibility but do not establish match correctness or localization accuracy.Source: F. Application on a Self-Developed ROV · p. 7
  • On collected lake images, UIQM increased by 24.43% and UCIQE by 29.57%, but no-reference scores do not independently establish fidelity to true scene colors.The findings support real-time enhancement on the sampled lake imagery.Source: F. Application on a Self-Developed ROV · p. 7

V. CONCLUSION

The conclusion presents PSD as a training framework for compressing high-performance underwater image enhancement models without adding inference-time modules. Across backbones and datasets, compact students improve; UDD detection and ROV deployment provide downstream and field evidence, while long-duration and closed-loop deployment remain future validation.

V. CONCLUSION

  • PSD combines degradation-aware weighting, reliability-gated low- and high-frequency transfer, and image-space compensation without adding inference-time modules.Long-duration and closed-loop deployment are left for future validation.Source: §V · pp. 8–9
  • Across three backbones and datasets, PSD consistently improved students retaining 0.94–4.6% of teacher parameters and 1.2–2.5% of teacher GMACs.The conclusion positions PSD as a way to compress high-performance underwater enhancement models for resource-constrained devices.Source: §V · pp. 8–9
  • UDD experiments showed benefits for downstream detection, with YOLOv9s detections shown for raw images and teacher/student outputs across three backbones.The detectors were trained and tested in their corresponding domains; the figure caption does not specify detection scores.Source: §V · p. 9, Figure 7 · p. 9
  • ROV deployment demonstrated real-time throughput and improved field-image quality, with speed and no-reference quality reported for Dushu Lake images.The ROV figure identifies onboard GoPro and stereo-camera locations; the supplied table captions do not give metric values.Source: §V · p. 9, Table V · p. 9, Table VI · p. 9
  • The ROV SIFT figure compares raw-frame matches with matches after NAFNet-student enhancement and labels retained-match counts for each pair.The caption identifies yellow raw-frame matches and green enhanced-frame matches but states no comparison outcome.Source: Figure 9 · p. 9
Loading 2609.34795v1…