Coarse-to-Fine Grained Alignment Video Super-Resolution for Underwater Camera

Abstract

Artificial intelligence advancements have notably improved consumer electronics, especially in portable mobile underwater cameras. Yet, the large size of underwater video files and mobile device data transfer limitations hinder this progress. Hence, there's a crucial need for super-resolution technology tailored to underwater video content. In this paper, given the scarcity of underwater scenes in existing benchmark video datasets, we have taken the initiative to construct the pioneering underwater video dataset UW206, which comprises 206 authentic underwater videos exhibiting diverse environments. Then we design the underwater video super-resolution (UWVSR) based on coarse and fine grained feature alignment. UWVSR initially processes consecutive multi-frame images, employs optical flow extraction to achieve a preliminary alignment of features in reference and adjacent frames, and subsequently refines alignment through a spatial attention-based mechanism to capture supplementary information. Finally, the image features facilitate video reconstruction through pixel shuffle upsampling. We qualitatively and quantitatively evaluate the UWVSR on the dataset UW206 and compare it to state-of-the-art methods. Experimental results demonstrate that UWVSR significantly enhances video quality in underwater super-resolution tasks.