arxiv: 2605.13720 · v1 · submitted 2026-05-13 · 📡 eess.IV

Recognition: unknown

An Underwater Dehazing Network with Implicit Transmission Estimation

Sahana Ray , Sanjay Ghosh

Authors on Pith no claims yet

Pith reviewed 2026-05-14 17:37 UTC · model grok-4.3

classification 📡 eess.IV

keywords underwater image enhancementdehazingBeer-Lambert lawimplicit depth estimationtransmission maplightweight networkimage restoration

0 comments

The pith

UDehaze-iT dehazes underwater images by implicitly estimating depth and deriving transmission maps from the Beer-Lambert law with learnable coefficients.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a hybrid network for restoring clarity to underwater images degraded by wavelength-dependent absorption and scattering. It estimates scene depth inside the network rather than from labels, then computes per-channel transmission using the Beer-Lambert law whose attenuation coefficients are learned during training. Atmospheric light is modeled as simple per-channel scalars and a lightweight residual block removes leftover artifacts. A composite loss combining pixel, frequency, reconstruction, and regularization terms trains the model to roughly 0.9 million parameters. The resulting system matches or approaches the performance of larger methods on the UIEB and UFO-120 benchmarks while retaining physical structure in its transmission estimates.

Core claim

UDehaze-iT estimates scene depth implicitly and forms per-channel transmission maps through the Beer-Lambert law whose attenuation coefficients are free parameters. Atmospheric light is recovered as a semi-classical per-channel scalar, after which a zero-initialized residual refiner corrects residual errors. Training uses a five-term composite loss that includes L1 fidelity, multi-scale patchwise DCT, forward-model reconstruction, and two regularizers. With approximately 0.9 million parameters the network reaches competitive scores on the UIEB and UFO-120 real-world underwater datasets.

What carries the argument

Implicit depth estimation that feeds a Beer-Lambert transmission model whose per-channel attenuation coefficients are learned as network parameters.

If this is right

Transmission maps produced by the network remain physically interpretable because they are generated from the Beer-Lambert relation rather than learned as arbitrary images.
The small parameter count supports deployment on autonomous underwater vehicles that must run in real time on limited hardware.
Training requires no paired depth supervision, lowering the cost of creating new underwater datasets.
The semi-classical atmospheric-light estimate and residual refiner allow the model to handle scenes where the basic physical model is only approximately valid.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same implicit-depth-plus-learnable-coefficient pattern could be adapted to other scattering media such as fog or turbid medical imaging by changing only the forward model inside the loss.
Because the attenuation coefficients are per-channel and learned, they may encode water-body-specific absorption spectra that could later be used for water-type classification.
The architecture's emphasis on a single forward-model reconstruction term suggests it could be extended to joint dehazing and depth estimation without additional supervision branches.

Load-bearing premise

Implicit depth estimation together with learnable per-channel attenuation coefficients is enough to capture the dominant scattering and absorption effects in real underwater scenes without explicit depth labels or more detailed physics.

What would settle it

A test set that supplies accurate depth maps for the same scenes; if the network's implicit depth deviates sharply from those maps and its visual quality falls below classical physics-based methods, the central modeling claim would be refuted.

Figures

Figures reproduced from arXiv: 2605.13720 by Sahana Ray, Sanjay Ghosh.

**Figure 2.** Figure 2: Visual comparison of underwater image enhancement results on representative samples from the UIEB dataset. Each row corresponds to a distinct [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

read the original abstract

Underwater images suffer from wavelength-dependent light absorption and scattering, which reduces visual quality. This phenomenon could limit the operational reliability of autonomous underwater vehicles, marine surveys, and offshore inspection systems. Purely classical methods often achieve suboptimal performance in real-world datasets, while purely data-driven methods lack physical interpretability. In this letter, we propose UDehaze-iT, a deep network for underwater image enhancement that estimates scene depth implicitly and derives per-channel transmission through the Beer-Lambert law with learnable attenuation coefficients. We estimate atmospheric light as a semi-classical per-channel scalar, and a zero-initialized residual refiner corrects remaining artefacts after dehazing. To effectively train our method, we apply a composite loss function consisting of five key terms: a L1 loss, a multi-scale patchwise DCT loss, a forward model reconstruction loss, and two regularization terms. With ~0.9M parameters, UDehaze-iT achieves competitive performance on UIEB and UFO-120 datasets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

UDehaze-iT pairs implicit depth with learnable Beer-Lambert transmission in a 0.9M-param net, but without depth checks or ablations the physics contribution stays unproven.

read the letter

This paper introduces UDehaze-iT, a compact network for underwater dehazing that estimates depth implicitly and feeds it into per-channel transmission via a learnable Beer-Lambert law, then applies semi-classical atmospheric light and a zero-initialized residual refiner trained on a five-term composite loss. It reports competitive results on UIEB and UFO-120 while staying under a million parameters. The specific architecture is new: the blend of implicit depth, learnable attenuation coefficients, and the refiner step gives a structured yet lightweight alternative to pure data-driven or pure classical approaches. The design keeps parameter count low and includes a forward-model reconstruction term in the loss, which is a sensible way to tie the output back to the input formation model. The zero-init refiner is a clean choice that lets the initial dehazing step do most of the work. The main limitation is the missing validation for the claimed physics path. The paper supplies no depth ground-truth comparison, no correlation between estimated and expected transmission, and no ablation that disables the transmission module while holding the refiner and loss fixed. As a result it is hard to tell whether the reported performance comes from the Beer-Lambert component or simply from the refiner and the multi-scale DCT plus regularization terms. The learnable coefficients also reduce the physical interpretability that the method advertises. This work is aimed at engineers building efficient enhancement pipelines for AUVs and marine surveys who want a model that is small and somewhat structured. If the full manuscript contains proper baselines, ablations, and some analysis of the depth maps, it deserves a serious referee to check the numbers and decide whether the physics modeling adds measurable value over simpler alternatives.

Referee Report

1 major / 1 minor

Summary. The manuscript presents UDehaze-iT, a lightweight (~0.9M parameter) deep network for underwater image dehazing. It implicitly estimates scene depth, derives per-channel transmission maps from the Beer-Lambert law using learnable attenuation coefficients, estimates atmospheric light as per-channel scalars, and applies a zero-initialized residual refiner. Training employs a composite loss with L1, multi-scale patchwise DCT, forward-model reconstruction, and two regularization terms. The central claim is competitive performance on the UIEB and UFO-120 datasets.

Significance. If the performance claims hold under proper quantitative evaluation, the work offers a parameter-efficient hybrid approach that injects a simple physical transmission model into a data-driven pipeline. This could be useful for real-time marine applications where model size and partial interpretability matter, provided the implicit depth component demonstrably contributes beyond the refiner and loss terms.

major comments (1)

[Method and Experiments] The headline claim that implicit depth estimation plus per-channel Beer-Lambert transmission yields competitive results requires evidence that the learned depth produces physically plausible transmission values. No depth ground-truth comparison, correlation metric, qualitative depth visualization, or ablation that isolates the transmission module (while freezing the refiner and loss) is supplied, so performance gains cannot be attributed to the claimed physics-informed path.

minor comments (1)

[Abstract] The abstract asserts 'competitive performance' on named datasets but supplies no numerical metrics, baseline comparisons, or pointers to result tables/figures, which hinders immediate assessment of the claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment point-by-point below and will revise the paper accordingly to provide stronger evidence for the physics-informed components.

read point-by-point responses

Referee: [Method and Experiments] The headline claim that implicit depth estimation plus per-channel Beer-Lambert transmission yields competitive results requires evidence that the learned depth produces physically plausible transmission values. No depth ground-truth comparison, correlation metric, qualitative depth visualization, or ablation that isolates the transmission module (while freezing the refiner and loss) is supplied, so performance gains cannot be attributed to the claimed physics-informed path.

Authors: We agree that additional validation is needed to link the implicit depth estimation to physically plausible transmission values and to isolate its contribution. The UIEB and UFO-120 datasets do not provide ground-truth depth, so quantitative depth comparisons or correlation metrics are not possible. In the revised manuscript we will add: (1) qualitative visualizations of estimated depth maps and derived per-channel transmission maps on representative images, demonstrating consistency with underwater optics (stronger red-channel attenuation); (2) a targeted ablation that replaces the Beer-Lambert transmission module with direct transmission regression while keeping the refiner and all loss terms frozen, reporting performance deltas on both datasets. These changes will allow readers to evaluate the contribution of the physics-informed path. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper proposes a trainable neural network (UDehaze-iT) that implicitly estimates depth inside the model and computes per-channel transmission via the Beer-Lambert law using learnable attenuation coefficients. These coefficients are optimized end-to-end during training rather than fixed external constants; the resulting transmission map is therefore an intermediate learned representation, not a first-principles prediction forced by construction from the input image alone. The central claim—competitive enhancement performance on the external UIEB and UFO-120 benchmarks with ~0.9 M parameters—is evaluated against held-out data and is therefore falsifiable. No self-citations, uniqueness theorems, or ansatzes imported from prior author work appear in the text. The architecture is a standard physics-inspired parameterization inside a data-driven pipeline and does not reduce any reported result to its own inputs by definition.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The method rests on the Beer-Lambert law as a domain assumption and introduces learnable attenuation coefficients plus loss weights as free parameters fitted to data; no new entities are postulated.

free parameters (2)

learnable attenuation coefficients
Per-channel scalars learned during training to parameterize transmission in the Beer-Lambert derivation.
composite loss weights
Relative weights for the five loss terms (L1, multi-scale DCT, reconstruction, and two regularizers) chosen to balance training.

axioms (1)

domain assumption Beer-Lambert law governs wavelength-dependent absorption and scattering in underwater media
Invoked to derive per-channel transmission from implicit depth and attenuation coefficients.

pith-pipeline@v0.9.0 · 5464 in / 1245 out tokens · 33703 ms · 2026-05-14T17:37:51.043390+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

[1]

Light in the sea,

S. Q. Duntley, “Light in the sea,”Journal of the Optical Society of America, vol. 53, no. 2, pp. 214–233, 1963

work page 1963
[2]

Recovery of underwater visibility and structure by polarization analysis,

Y . Y . Schechner and N. Karpel, “Recovery of underwater visibility and structure by polarization analysis,”IEEE Journal of Oceanic Engineer- ing, vol. 30, no. 3, pp. 570–587, 2006

work page 2006
[3]

Single image haze removal using dark channel prior,

K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 12, pp. 2341–2353, 2010

work page 2010
[4]

A revised underwater image formation model,

D. Akkaynak and T. Treibitz, “A revised underwater image formation model,”Proc. IEEE Conference on Computer Vision and Pattern Recog- nition, pp. 6723–6732, 2018

work page 2018
[5]

Single underwater image enhancement using depth estimation based on blurriness,

Y .-T. Peng, X. Zhao, and P. C. Cosman, “Single underwater image enhancement using depth estimation based on blurriness,”Proc. IEEE International Conference on Image Processing, pp. 4952–4956, 2015

work page 2015
[6]

Underwater image restoration based on image blurriness and light absorption,

Y .-T. Peng and P. C. Cosman, “Underwater image restoration based on image blurriness and light absorption,”IEEE Transactions on Image Processing, vol. 26, no. 4, pp. 1579–1594, 2017

work page 2017
[7]

Underwater image enhancement via piecewise color correction and dual prior optimized contrast enhancement,

W. Zhang, S. Jin, P. Zhuang, Z. Liang, and C. Li, “Underwater image enhancement via piecewise color correction and dual prior optimized contrast enhancement,”IEEE Signal Processing Letters, vol. 30, pp. 229–233, 2023

work page 2023
[8]

Lightness and retinex theory,

E. H. Land and J. J. McCann, “Lightness and retinex theory,”Journal of the Optical Society of America, vol. 61, no. 1, pp. 1–11, 1971

work page 1971
[9]

Fast scale- adaptive bilateral texture smoothing,

S. Ghosh, R. G. Gavaskar, D. Panda, and K. N. Chaudhury, “Fast scale- adaptive bilateral texture smoothing,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 30, no. 7, pp. 2015–2026, 2019

work page 2015
[10]

Underwater image enhance- ment via extended multi-scale retinex,

S. Zhang, T. Wang, J. Dong, and H. Yu, “Underwater image enhance- ment via extended multi-scale retinex,”Neurocomputing, vol. 245, pp. 1–9, 2017

work page 2017
[11]

On the duality between retinex and image dehazing,

A. Galdran, A. Alvarez-Gila, A. Bria, J. Vazquez-Corral, and M. Bertalm ´ıo, “On the duality between retinex and image dehazing,” Proc. IEEE Conference on Computer Vision and Pattern Recognition, pp. 8212–8221, 2018

work page 2018
[12]

Underwater optical image contrast enhancement via color channel matching,

X. Chen, S. Sun, Y . Gao, W. Zhao, and W. Zhang, “Underwater optical image contrast enhancement via color channel matching,”IEEE Geoscience and Remote Sensing Letters, 2025

work page 2025
[13]

An underwater image enhancement benchmark dataset and beyond,

C. Li, C. Guo, W. Ren, R. Cong, J. Hou, S. Kwong, and D. Tao, “An underwater image enhancement benchmark dataset and beyond,”IEEE Transactions on Image Processing, vol. 29, pp. 4376–4389, 2019

work page 2019
[14]

Underwater im- age enhancement via learning water type desensitized representations,

Z. Fu, X. Lin, W. Wang, Y . Huang, and X. Ding, “Underwater im- age enhancement via learning water type desensitized representations,” Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2764–2768, 2022

work page 2022
[15]

An efficient approach for underwater image improvement: Deblurring, dehazing, and color correction,

A. R. Espinosa, D. McIntosh, and A. B. Albu, “An efficient approach for underwater image improvement: Deblurring, dehazing, and color correction,”Proc. IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 206–215, 2023

work page 2023
[16]

The beer-lambert law,

D. F. Swinehart, “The beer-lambert law,”Journal of chemical education, vol. 39, no. 7, p. 333, 1962

work page 1962
[17]

Shallow-UWnet: Compressed model for underwater image enhancement (student abstract),

A. Naik, A. Swarnakar, and K. Mittal, “Shallow-UWnet: Compressed model for underwater image enhancement (student abstract),” vol. 35, no. 18, pp. 15 853–15 854, 2021

work page 2021
[18]

Fast underwater image enhancement for improved visual perception,

M. J. Islam, Y . Xia, and J. Sattar, “Fast underwater image enhancement for improved visual perception,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3227–3234, 2020

work page 2020
[19]

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

work page 2004
[20]

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,”Proc. IEEE Conference on Computer Vision and Pattern Recognition, pp. 586– 595, 2018

work page 2018
[21]

Underwater image enhancement via medium transmission-guided multi-color space embedding,

C. Li, S. Anwar, J. Hou, R. Cong, C. Guo, and W. Ren, “Underwater image enhancement via medium transmission-guided multi-color space embedding,”IEEE Transactions on Image Processing, vol. 30, pp. 4985– 5000, 2021

work page 2021
[22]

Underwater image enhancement based on deep learning and image formation model,

X. Chen, P. Zhang, L. Quan, C. Yi, and C. Lu, “Underwater image enhancement based on deep learning and image formation model,”arXiv preprint arXiv:2101.00991, 2021

work page arXiv 2021
[23]

Underwater Ranker: Learn which is better and how to be better,

C. Guo, R. Wu, X. Jin, L. Han, W. Zhang, Z. Chai, and C. Li, “Underwater Ranker: Learn which is better and how to be better,” in Proceedings of the AAAI conference on artificial intelligence, vol. 37, no. 1, 2023, pp. 702–709

work page 2023