Transmission and depth estimation for physics-based underwater image enhancement

Water absorbs and scatters light in a wavelength-dependent way, and the result for underwater imagery is a predictable set of problems: color cast, washed-out contrast, and degraded structure. Physics++ addresses these through a deterministic pipeline built from closed-form stages: depth-aware attenuation modeling, edge-preserving guided-filter transmission refinement, and a residual-constrained fusion step that blends a direct radiance estimate against a dark-channel-guided prior. None of these stages involve learned parameters, so identical input always produces identical output and no training data is required. We evaluated the method on four public benchmarks (LSUI, UIEB, EUVP, and SQUID) against seven other physics-based configurations. Physics++ posted the best SSIM on LSUI (0.7944), the best PSNR (17.55 dB) and UCIQE (52.89) on UIEB, and the best average rank across all four datasets (3.00 of 8); the gains over the UDCP and WCID baselines were statistically significant ( p < 0.001), with large effect sizes on UCIQE and UIQM. We do not obscure the method's ceiling: its absolute PSNR sits 3–8 dB below current learning-based state of the art, and we position it deliberately as a training-free, interpretable alternative rather than a fidelity-maximizing competitor. To test whether that ceiling can be narrowed without abandoning the physics, we also trained a compact hybrid variant, Physics++-Hybrid, conditioning a learned network on the same transmission signal; on a held-out EUVP split it reaches 23.35 dB mean PSNR, an encouraging result, though measured under a different protocol and not directly comparable to the core result above.