视频背景替换技术研究与应用

【作者】 赵健；

【导师】 项欣光；

【作者基本信息】 南京理工大学 ， 计算机技术（专业学位）， 2017， 硕士

【摘要】 视频背景替换是指从一段视频序列中提取出前景并合成到新背景的过程,它在视频编辑领域中发挥着重要作用,具有很高的应用价值。视频背景替换的主要实现方式是视频抠图,然而抠图一直以来是一个欠约束的问题,必须要施加一些限制条件来协助解决,早期的方法通过限制背景颜色或者在特定环境下拍摄来抠图,这极大影响了抠图技术的发展。因此,为了突破这一限制,近几年兴起了适用范围更广的自然图像抠图技术的研究。视频抠图作为图像抠图的扩展,它的难点在于视频的数据量很大,由于大部分图像抠图方法都需要一定用户交互操作,将图像抠图的方法直接应用到视频上是不现实的。因此,如何减少用户交互、改善抠图质量、提高抠图速度是我们当前研究的重点。本文通过研究现有的图像、视频抠图技术,提出并实现了两种不同类型的视频抠图方法。主要研究内容如下:1)提出了一种基于运动估计的视频抠图方法。首先,该方法不使用三分图作为输入,而是将选取的关键帧粗略分割成前景和背景。然后,使用基于块匹配的运动估计算法将关键帧的分割结果传播到非关键帧。最后,使用改进的引导滤波器对视频分割序列做优化并最终替换新的背景。2)提出了一种结合贝叶斯估计和KNN抠图的视频背景替换方法。首先,该方法只需用户提供视频首帧的不透明度。然后,利用贝叶斯估计自动生成其余各帧的三分图,再通过KNN抠图算法计算出每一帧的不透明度。最后,用改进的引导滤波器优化抠图结果,并为视频序列替换新的背景图片。3)设计并实现了一种交互式的背景替换系统,包含图像与视频抠图的基本功能。更多还原

【Abstract】 Video background replacement refers to the process of extracting foreground from a video sequence and composite it with a new background,it plays an important role in video editing field and has very high application value.Video matting is the main way to realize video background replacement.However,matting has been an under-constraint problem for a long time,it can’t be solved appropriately without some constraints.Early methods either fix the background or photograph in a specific environment to help matting,which greatly affect the development of matting technology.So,in order to break through this limitation,many researchers begin to study natural image matting technology,which is much more adaptive in many fields.Video matting,extended from image matting,is limited by its data size.Since most image matting methods require some user interaction,we can’t apply them to videos directly.How to reduce the amount of user interaction,how to improve the matting quality and how to speed up the matting process are key points of our current study.In this paper,we focus on the study of existing image and video matting methods,and propose two different methods for video matting.The main work of our study is as follow:1)We propose a video matting approach based on motion estimation.Firstly,trimaps are not required in our method,we select some keyframes and segment them into rough foreground and background.Then,we propagate segmentation results of keyframes to other frames by a motion estimation algorithm based on block matching.Finally,we use the improved guided filter to refine the result of the segmentation sequence and replace the background of original video.2)We propose a video matting approach based on Bayesian estimation and KNN matting.Firstly,we only require the user to provide the alpha of first frame,this work can be done by any image matting method.Then,we use Bayesian estimation to automatically produce the trimap of other frames,and apply KNN matting algorithm to compute the alpha of each frame.Finally,we use the improved guided filter to refine the result,and composite the alpha sequence with a new background.3)We design and realize an interactive background replacement system which contains the basic functions of image and video matting.更多还原