DUVR-Net: A Dual-Path Underwater Video Restoration Network Based on Middle-Frame Residual and Transformation Convolution Kernel

Underwater video enhancement aims to improve the visibility and frame quality of underwater videos, which is crucial for marine and lake research and exploration. However, existing methods primarily rely on image enhancement algorithms that independently enhance individual frames, neglecting the frame-to-frame continuity and style similarity relationships inherent in underwater videos. Moreover, current methods require excessive parameters and slow training processes to accommodate 4K video enhancement. To address these challenges, we propose the Dual-path Underwater Video Restoration Network(DUVR-Net), which first enhances the middle frame and then transfers the enhancement effect to adjacent frames through a dual-path approach using residual compensation and transformed convolution kernels for inter-frame information transmission.Without the need to enhance each frame independently,our model requires only approximately 90k parameters. Experiments on real underwater videos demonstrate that our network can effectively accomplish 4K underwater video enhancement with impressive results.