UVENet: A novel end-to-end model for temporal consistency in underwater video enhancement.

Abstract

Deep-sea submersibles collect extensive underwater data, supporting the study and exploration of marine ecosystems. Efficient processing and analysis of underwater images and videos are crucial for applications like ecological monitoring, resource assessment, and environmental conservation. These underwater visual tasks face significant challenges due to the effects of wavelength-dependent absorption and scattering, which lead to issues like color casts, blurred details, and low contrast. Despite significant advancements in underwater image enhancement (UIE), underwater video enhancement (UVE) remains underdeveloped. Traditional UVE methods that extend UIE techniques by independently enhancing each frame fail to address temporal consistency, leading to artifacts such as flickering. Constructing high-quality paired datasets for UVE is also a significant challenge, as obtaining both real underwater videos and corresponding ground truth videos is often impractical due to environmental limitations. To address these challenges, we introduce UVENet, a novel end-to-end UVE model that leverages multi-frame inputs and incorporates Feature Alignment and Aggregation Modules (FAAMs) to ensure effective spatial alignment and feature aggregation, thereby preserving temporal consistency. To support the training and evaluation of UVENet, we construct the first synthetic underwater video enhancement dataset (SUVE), which consists of 840 pairs of videos generated using underwater neural rendering (UWNR) technology. Extensive experiments on both synthetic and real underwater videos validate the effectiveness of our approach. Our code is available at https://github.com/ddz16/UVENet.