ReX-Net: A reflectance-guided underwater image enhancement network for extreme scenarios

Abstract

Due to the complex underwater environment, underwater images exhibit different degradation characteristics, severely affecting their practical applications. Although underwater image enhancement networks with physical priors exist, the statistical priors are not applicable in extreme underwater scenes. Therefore, we propose ReX-Net, a reflectance-guided underwater image enhancement network for extreme scenarios. ReX-Net leverages the complementary information of reflectance and the underwater image obtained through the original encoder and reflectance encoder to minimize the impact of different scene environments. As underwater images contain object information at different scales, the encoder includes a TriFuse reflected-image object extractor module (TRIOE), which employs Tri-scale convolutions to capture features at different scales and utilize attention mechanisms to enhance channel and spatial information. In the decoder, we design a context-sensitive multi-level integration module (CSMLI) to fuse feature vectors at different resolutions, thereby improving the expressiveness and robustness of features while avoiding artifacts and ensuring pixel accuracy. Experiments on multiple datasets demonstrate that ReX-Net outperforms existing methods. Furthermore, application experiments show the practicality of ReX-Net in other visualization tasks.