Abstract

Underwater imaging inherently suffers from severe quality degradation due to physical phenomena including light absorption, wavelength-dependent color attenuation, and severe backscattering. These optical distortions compromise object boundaries and feature distinctiveness, presenting substantial hurdles for downstream computer vision tasks such as instance segmentation. In this paper, we conduct an extensive empirical study investigating the impact of diverse underwater image enhancement (UIE) techniques on the accuracy and robustness of deep learning-based instance segmentation models. By evaluating both traditional physical model-based restoration techniques and modern deep learning generative models across varied aquatic datasets, we identify critical trade-offs between perceptual enhancement metrics and segmentation precision. Our findings offer practical guidelines for designing unified visual perception pipelines in aquatic robotics and marine observation.