ColorRefineNet: Color Refining Underwater Image Enhancement based on Attention Network

In recent times, the challenges associated with underwater images have gained more attention, particularly in color enhancement, which are significant obstacles for high-level computer vision tasks. To solve these issues, this paper proposes a color refinement network, named ColorRefineNet, which is designed to enhance the brightness of color and contrast of underwater images while preserving structure details, sharpness, and colorfulness. In ColorRefineNet, a channel and spatial attention-based Convolutional Block Attention Module (CBAM) attention module is incorporated into each ConvBlock of the baseline Shollow-UWnet. Additionally, a structural similarity (SSIM) loss function is integrated as an additional loss to enhance the color and contrast of underwater images. Comprehensive experiments on eight underwater image datasets, such as EUVP-Dark, UIEB, UFO-120, U45, C60, T90, ColorChecker, and OceanDark, show that ColorRefineNet outperforms UCIQE scores than the baseline Shallow-UWnet and state-of-the-art (SOTA) methods.