Fast Adaptive Self-Supervised Underwater Image Enhancement.

Mengxiao Huang,Yang Wang,Weiqi Zou,Yang Cao

DOI: https://doi.org/10.1109/icip46576.2022.9897298

2022-01-01

Abstract:Wavelength-dependent light absorption and scattering result in the color cast and contrast degradation, which degrades the visibility of underwater images. Deep-learning-based underwater image enhancement has made significant progress in recent years. However, existing algorithms rely heavily on massive data for training, and the difficulty of collecting massive data in real-world environments limits their applicability. To alleviate this problem, we propose a fast adaptive self-supervised underwater image enhancement method, which models the entire water domain in terms of color and contrast distribution with only a few images. Specifically, to learn the mapping of color correction and contrast enhancement, our framework contains two sub-networks: a Color Mapping Net (CLM-Net) and a Contrast Mapping Net (CTM-Net). The CLM-Net learns the color correction mapping under the constraint of gray world assumption in a self-supervised manner. Besides, the CTM-Net estimates the contrast enhancement mapping under the constraint of the dark channel prior. Thanks to adaptive self-supervised training, we can primarily alleviate the need for collecting massive data and achieve effective underwater image enhancement with only ten images. Experiments demonstrate that the proposed algorithm achieves state-of-the-art performance on four benchmark datasets.