Underwater video consistent enhancement: a real-world dataset and solution with progressive quality learning

Underwater image and video enhancement task is of great significance for ocean exploration. Compared with a single image, underwater video enhancement is more susceptible to scene and light changes, which causes inconsistency between the frames of enhanced video. In this paper, we construct a high-quality dataset named UVE38K to establish a benchmark for underwater video enhancement, which consists of 50 real-world videos from various environments. To understand the difference in quality between underwater video frames. We propose a quality superiority decision network (QSDNet) to distinguish the high-quality and low-quality frames in enhanced videos. Our QSDNet can achieve an accuracy of 87.9%. We also propose two underwater video enhancement algorithms PUVE and BUVE for online and offline situations respectively. Experiments on the UVE38K dataset show that our methods outperform existing methods.