视频帧插值篡改的被动检测综述  

A Survey of Blind Detection for Video Frame Interpolation

作　　者：杨高波 YANG Gaobo(College of Computer Science and Electronic Engineering,Hunan University,Changsha 410082,Hunan)

机构地区：[1]湖南大学信息科学与工程学院,湖南长沙410082

出　　处：《四川师范大学学报（自然科学版）》2023年第4期438-447,共10页Journal of Sichuan Normal University（Natural Science）

基　　金：国家自然科学基金(61970142);湖南省自然科学基金面上项目(2020JJ4029);信息安全国家重点实验室开放课题(2021-ZD-07)。

摘　　要：视频帧插值是指利用运动补偿插值或者基于深度学习的视频帧合成并且插入到视频,属于视频特有的时域操作.传统的视频帧插值和深度视频帧插值都可以用于提升视频帧率上转,也可能用于提升虚假的视频码率,以吸引用户访问.系统地综述视频帧插值的被动检测技术.首先,分析视频帧插值技术的帧合成机理,尤其突出深度视频帧插值技术,包括其采用的深度网络模型及其特点,以挖掘遗留的细微痕迹.然后,从篡改手段和遗留痕迹的角度,总结基于手工特征和基于深度学习的视频帧插值被动检测的典型方法,并且从采用的特征和优缺点的角度进行比较.最后,指出视频帧插值篡改检测领域在深度学习时代存在的技术挑战和未来的发展趋势.Video frame interpolation refers to motion compensation interpolation or deep learning-based video frame synthesis for the generation of intermediate frame to be interpolated,which is a temporal manipulation specific to digital video.The conventional and deep video frame interpolation can be used for video frame rate up-conversion,and might also be used for promoting fake video bitrate to attract user’s visit.This work systematically surveys the blind detection techniques for video frame interpolation.Firstly,we analyze the frame synthesis mechanism for video frame interpolation,especially deep video frame interpolation,including deep models and their characteristics,to find subtle tampering traces.Secondly,we summarize the typical hand-crafted features and deep learning-based blind detection methods from tampering strategy and left traces,and make comparisons among them in terms of their features,advantages and disadvantages.Finally,we point out the technical challenges that still exist in the passive forensics field of video frame interpolation and future works.

关 键 词：视频帧插值 数字视频取证 被动取证 真实性鉴别 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]