具有自动场景划分的多模板视频着色  

Multi Template Video Coloring With Automatic Scene Division

作　　者：刘兴业 高媛[1] 秦品乐[1] 赵一飞 LIU Xingye;GAO Yuan;QIN Pinle;ZHAO Yifei(School of Data Science and Technology,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学大数据学院,山西太原030051

出　　处：《中北大学学报（自然科学版）》2023年第4期388-396,共9页Journal of North University of China(Natural Science Edition)

基　　金：山西省“揭榜挂帅”重大专项(202101010101018)。

摘　　要：针对目前视频着色的方法存在视频转换场景时颜色发生明显偏移和时间不一致性的问题,提出了一种具有自动场景划分的多模板视频着色方法。该方法根据视频场景的不同,将整个视频分成若干组,每组利用图像检索的方式找到一个合适的参考图像,使用参考图像与前一帧图像共同指导视频着色,同时,计算余弦相似性对各个边界帧进行比较,保证具有相似场景的视频使用相同的参考图像,使得整个视频满足时间一致性。实验结果表明,所提出的方法可以有效解决视频着色中时间不一致性与颜色抖动的问题,并且在多个场景下的视频着色效果都有明显提升,与全自动视频着色、单模板视频着色和其他多模板视频着色相比,峰值信噪比(PSNR)分别提升了6.01%,3.81%和5.02%,且产生的视频颜色更加自然。Aiming at the problems of obvious color shift and time inconsistency in current video coloring methods,a multi template video coloring method with automatic scene division was proposed.According to the different video scenes,the whole video was divided into several groups.Each group used image retrieval to find a suitable reference image,used the reference image to guide the video coloring together with the previous frame,and calculated the cosine similarity to compare each boundary frame to ensure that the videos with similar scenes used the same reference image,so that the whole video met the time consistency.The experimental results show that the proposed method can effectively solve the problems of time inconsistency and color jitter in video coloring,and the video coloring effect in multiple scenes is significantly improved.Compared with automatic video coloring,single template video coloring and other multi template video coloring,the peak signal-to-noise ratio(PSNR)is increased by 6.01%,3.81%and 5.02%respectively,and the resulting video color is more natural.

关 键 词：视频着色 深度学习 图像检索 时间一致性 镜头边界检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]