Two-stage underwater image enhancement using domain adaptation and interlacing transformer

Abstract

Underwater images typically suffer from quality defects such as impurities, light scattering, absorption, color casts, etc. Various underwater image enhancement (UIE) techniques based on deep learning have recently been proposed and have achieved remarkable results, but they still have qualitative difficulty from irregular color channel loss. This problem is usually caused by the difficulty of distinguishing between the color cast of the underwater and the irregular scattering of the object. In this work, we propose a two-stage approach to refine domain information and semantic information, respectively: domain adaptation network and image enhancement network. In the domain adaptation network, we introduce Sobel sparse attention, which can separate the semantic information of underwater images and remove domain information. A U-shape architecture of the transformer is designed for efficient decoding of the image enhancement network. In addition, the interlacing position embedding is adopted, which can solve the scattering and blurring of light that occurs locally. To validate our proposed method, we conducted various experiments using non-reference, full-reference, and synthetic underwater images. The experimental results demonstrate that our method outperforms the state-of-the-art qualitatively and quantitatively.