Abstract:Underwater imaging faces severe color distortion and contrast degradation due to light absorption and scattering. Prevailing deep learning methods require extensive paired data and suffer from complex architectures. This paper proposes a zero-reference learning framework using parameterized curve-correction modeling, reformulating enhancement as dual-stage parameter map estimation. In the color enhancement stage, a physically grounded color correction model derived from haze imaging principles is developed. A lightweight module estimates dynamic adjustment parameters while optimizing pixel-level dynamic range. In the luminance optimization stage, we deploy a non-linear luminance mapping function coupled with an adaptive brightness correction model. This rectifies residual dynamic range compression and illumination constraints from prior processing, performing global luminance calibration and model bias correction. Critically, the framework operates without paired underwater data. Benchmark evaluations demonstrate competitive performance in both subjective and quantitative metrics, validating the methodological efficacy and practical viability of this physics-inspired correction framework for underwater vision.
