Shot-Again: Self-Supervised Underwater Image Enhancement With Transmission-Space Regularization and Structure-Channel Feature Regulation

Abstract

Underwater image enhancement remains a challenging ill-posed problem because wavelength-dependent attenuation, scattering, and color distortion permit multiple plausible restoration solutions. Although self-supervised approaches reduce the dependence on paired underwater data, reconstruction-based objectives alone may leave the latent transmission estimation under-constrained. In this work, we propose Shot-Again, a self-supervised underwater image enhancement framework that extends re-degradation-based learning with an additional transmission-space regularization signal. Specifically, the controlled proxy re-degradation operation induces a relative relationship between the transmission estimates of the original and proxy observations, which is exploited as a soft regularization cue rather than being treated as an exact physical law for a newly captured scene. We further introduce a Structure-Preserving Block (SPB) and Spatial Group Cross-Channel Attention (SGCA) to provide complementary structure- and channel-aware feature regulation. Experiments across multiple underwater benchmarks, together with targeted architectural and low-light diagnostics, demonstrate competitive self-supervised restoration performance and a favorable quality–efficiency trade-off under diverse degradation conditions. The expanded comparison also reveals metric-dependent strengths among different enhancement paradigms rather than universal superiority of a single method. The implementation is available at: https://github.com/FengSY-JLU/Shot-again/tree/main