A Two-stage Progressive Network for Underwater Image Enhancement
Lingbo Kong, Yankai Feng, Shuxin Yang, Xu Gao · 2024
Due to the selective absorption and random scattering of light by the water medium, underwater images often suffer from color deviation and low contrast. To solve the above two degradation problems, a two-stage progressive network (TSPNet) for underwater image enhancement is proposed in this paper. More specifically, in the first stage of the network, we propose a channel-separated color restoration module (CSCRM), which takes into account the physical property of uneven attenuation of light underwater, and accomplishes the independent adjustment of the three color channel pixels separately through a divide-and-conquer strategy. In the second stage of the network, we further propose the multi-scale edge sharpening module (MSESM), which establishes mapping relationships at multiple scales to enhance different levels of underwater target details, and the embedded dense residual block (DRB) further enhances the adaptive ability of the model. In order to reveal the advancement of the proposed model, we do a detailed comparison with five different types of enhancement methods on a benchmark dataset, and the qualitative and quantitative results validate the superiority of the model to enhance the quality of underwater images.