基于深度学习的水面深度视频稳像  

Deep Learning-Based Video Stabilization for Water Surface Scenes

作　　者：杨礼华 陈泽钰 施俊杰 潘海朗[1,2] 杨劲松 LIU Jianguo YANG Lihua;CHEN Zeyu;SHI Junjie;PAN Hailang;YANG Jingsong;LIU Jianguo(School of Electronic and Optical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;State Key Laboratory of Satellite Ocean Environment Dynamics,Second Institute of Oceanography,Ministry of Natural Resources,Hangzhou 310012,China;Department of Earth Science and Engineering,Imperial College London,London SW72AZ,United Kingdom)

机构地区：[1]南京理工大学电光学院,江苏南京210094 [2]自然资源部第二海洋研究所卫星海洋环境动力学国家重点实验室,浙江杭州310012 [3]帝国理工学院地球科学工程系,英国伦敦SW72AZ

出　　处：《移动通信》2024年第11期86-91,129,共7页Mobile Communications

基　　金：国家自然科学基金“近岸条件下北斗导航反射信号海面风速反演能力增强机制研究”(42306200);国家自然科学基金青年科学基金项目“基于深度学习的热带气旋风场重构及风暴潮模拟”(42306216);国家重点研发计划项目“环境遥感巡查与立体智能感知”(2022YFC3103101)。

摘　　要：水面视频的稳像与深度稳像技术,在海洋监测、渔业资源调查、水上运动赛事直播等多个领域均展现出重要的应用价值。将深度学习运用于水面场景的深度视频稳像。首先使用软件Adobe Premiere Pro(Pr)的变形稳定器对水面视频进行稳像,以消除俯视角水面视频的无人机水平晃动。用裁剪率、失真值、累计光流等指标进行评估,结果显示稳像后视频质量损失较小,前后帧差异相较于原视频均在减少,因此稳像效果很好,视频稳定性得到了改善。然后采用一种神经视频深度稳定方法对水面进行深度视频稳像,该视频处理框架包括了深度预测器DPT-L和稳定网络两部分。实验结果表明,深度稳像后水面波浪与背景的边界变得更清楚,纹理变得清晰,其边界轮廓线与原视频的基本吻合,得到了很好的深度稳像效果。Video stabilization and deep video stabilization for water surface scenes have significant applications in fields such as ocean monitoring,fisheries resource surveys,and live streaming of water sports.This study applies deep learning to achieve deep video stabilization for water surface scenes.Initially,Adobe Premiere Pro’s Warp Stabilizer was used to stabilize water surface videos,reducing horizontal drone shake in top-down water surface footage.Metrics such as cropping ratio,distortion value,and cumulative optical flow were employed for evaluation.Results showed minimal quality loss post-stabilization,with reduced inter-frame differences compared to the original video,indicating improved video stability.Subsequently,a neural deep video stabilization method was applied,incorporating a depth predictor and a stabilization network.Experimental results demonstrate that after deep stabilization,boundaries between water waves and the background became sharper,textures were clearer,and boundary contours closely matched those of the original video,achieving excellent deep stabilization performance.

关 键 词：视频稳像 深度视频稳像 稳定性评估 水面场景 

分 类 号：TN929.5[电子电信—通信与信息系统]