Underwater Color Restoration with Vanishing Uncertainty
Grigory Solomatov, Derya Akkaynak
eess.IV, cs.CV
2026-08-16
Researchers at the University of Haifa, home of SeaThru, derive sufficient conditions under which underwater color restoration becomes provably solvable: when pixels in a uniform-color segment cover enough distinct distances, uncertainty intervals shrink to zero as camera resolution grows, with no assumptions on the camera.
Color distortion in underwater photographs comes from absorption and scattering: light attenuates per meter of water, wavelength by wavelength, and a blue-green veil accumulates with distance. Coral bleaching monitoring, chlorophyll estimation, and species classification all need color as a trustworthy signal. Physics-based methods like SeaThru (2019, from the same senior author, Akkaynak) produce convincing results, but validation is almost entirely empirical: run end-to-end on a test set with ground truth and hope deployment conditions match. Water's optical properties vary sharply in space and time, so that hope is thin.
The paper first proves how bad the general case is (Theorem III.4): given only pixel intensities and distances, the solution space is the entire real line. Any output is consistent with the input. Empirical validation has been the only option precisely because nobody knew which constraints suffice or whether real data satisfies them. This paper supplies the first sufficient conditions on the theory side.
The image formation model is the simplified radiative transfer equation: pixel value equals attenuated object radiance plus distance-accumulated backscatter, integrated over the visible spectrum. The central object is inherent radiance segmentation, partitioning the image into segments that share the same true color once water is removed. Both results revolve around it.
Result 1 (Theorem V.3): given segmentation, restoration reduces to function extrapolation with a computable error bound. Within each uniform-color segment, subtract the backscatter first; restoring to any distance z becomes extrapolating a decaying curve. Pick coefficients αₖ so that the linear combination ΣαₖP̄ₖ best matches the target. The only extra inputs: upper and lower bounds on the attenuation coefficient, one "infinite-distance" pixel for the pure-water backscatter, and the segmentation itself. The error term Ψ shrinks to zero as the number of distinct distances in the segment grows, via the Müntz–Szász theorem from approximation theory; the condition is Σzᵢ⁻¹ → ∞, meaning distances must not grow too fast. Intuition: pixels on the same fish, covering enough different distances, pin down the attenuation curve.
Result 2 (Theorem VI.8): under idealized conditions, the segmentation itself is algorithmically determinable. Three conditions on the continuous image: connected, each uniform-color region is a single connected patch; discrete, every region is sampled infinitely often; ε-separable, distinct inherent radiances stay at least ε apart in the distance-brightness space, which strongly forbids metamerism. Add a continuous depth map, and an algorithm exists that groups pixels correctly at sufficiently high resolution.
Two improvements over the authors' own prior preprint (arXiv 2605.03837, May 2026): no assumptions on the camera whatsoever, monochrome cameras qualify and spectral sensitivity functions need not be known; and segmentation goes from oracle-given to conditionally determinable.
Pure theory: no experiments on real datasets and no metric tables. The quantitative content is worked examples and error bounds:
| Item | Result | ||
| Solution space in the general setting (Thm III.4) | All of R | ||
| Error with segmentation | ±Ψ·∫ | L−B | S dλ, with Ψ→0 as distinct distances →∞ |
| Camera assumptions | None; monochrome supported, spectral sensitivity unknown | ||
| Fig. 2 example | Uniform sampling over z=10–15; extrapolated curve hugs the true curve with explicit bounds |
Figure 2 also plots a log-linear regression curve for comparison (fit a line against log-distance and extrapolate, a common engineering shortcut): it visibly drifts off the true curve. The paper offers no quantitative comparison for this, so "the common shortcut drifts" stays at the level of a visual observation here.
For anyone in underwater imaging, this is foundation work, not a new tool. Neural methods for underwater restoration are numerous, and they share one gap: no theory says when the input information suffices for the output. These conditions now work as a checklist — when a new method claims to solve restoration, ask which constraints it implicitly relies on and whether real data satisfies them. Same logic as proving a benchmark is solvable before arguing about which model scores best.
One broadly useful byproduct is the zero camera assumptions. The prior preprint required a perfect hyperspectral camera; now any camera, monochrome included, qualifies in principle, with no spectral calibration. That moves the applicable range from the lab to arbitrary existing hardware.
Akkaynak is the first author of both SeaThru and the 2018 revised underwater image formation model this paper builds on. Read it as the theoretical footing her empirical line has lacked for a decade.
The authors are blunt that the conditions are "idealized beyond immediate applicability to real-world images":
My own doubts: the infinite-distance pixel may not exist in a real frame, and the paper does not discuss how to obtain one (treating a distant water column as infinity, say). The model also holds only for horizontal viewing, a hard constraint for the downward- and upward-looking cameras common in fieldwork; it gets one sentence in the derivation.