- 學位論文
藉由動作植入與影片修補實現影片重製
Video Reprocessing via Motion Interpolation and Video Inpainting
摘要
在影片中物件的行為是可以被更換的，亦即所謂的影片重製，雖然這項研究議題或許會引發一些較為負面的影響像是刻意將影片造假等，但是將這種技術應用於好的地方是會帶來相當大的貢獻，就有如成功的降低電影特效製造的成本或是讓人們方便編輯自己美好的回憶。而為了達到上述的目的，則必須實做出像是物件追蹤、動量估測、影像影片修補、動作分析與植入等技術。 在本文中，首先我們提出了一個經改良後的影像修補方法。這個影像修補方法可以適用於擁有各種不同特性的影片，像是屬於靜態背景的影片、內含有動態背景及物件有自身運動模式的等片等，均可使用我們提出的方法進行物件移除以及影像修補，修補完的結果亦是相當完善。其次，我們提出了一個全新的物件再造技術。這個技術納入了像是Mean Shift Segmentation顏色分群演算法、四步搜尋及十字菱形六角形搜尋等動量估測的方法以及在這次計畫中我們自己設計的動作分析與動作植入演算法。首先我們使用Mean Shift Segmentation將影片中的色彩做分群，以利統計出背景與物件各自得色彩分佈資訊，接著我們除了使用四步搜尋及十字菱形六角形搜尋演算法分別針對背景與物件實測出其內含的動量資訊以用於動作分析與動作植入程序，也使用了細化演算法來取得物件的骨架，並以此骨架為基本模組來推測出物件的新動作。在推出物件的新動作後，配合透過我們提出的影像修補方法所產生出的影片背景圖即可達成所謂影片重製的目標。
並列摘要
The behavior of people in a video can be altered. In order to change the content of video, issues such as object tracking, motion interpolation, video inpainting, and video layer fusing need to be implemented. In this dissertation, at first we extend an exemplar-based image inpainting algorithm by incorporating an improved patch matching strategy for video inpainting. The proposed new video inpainting algorithm produces very few “ghost shadows”, which were produced by most image inpainting algorithms directly applied on video. Secondly, we propose a novel motion interpolation algorithm by using the mean-shift segmentation and motion analysis technique. Mean shift segmentation is frequently used to extract objects from video according to its efficiency and robustness of non-rigid object tracking. For diminishing the computational complexity in motion estimation and object tracking process, several efficient block matching algorithms was used. In the motion analysis procedure, the stick figure of object obtained by thinning process is considered as guidance to gather the statistics of motion information. The model of new behavior of object is produced by motion analysis and used as an input to our motion interpolation procedure. In conclusion, a new video with different plots can be generated after the new object and corresponding background information are produced by our motion interpolation procedure and video inpainting technique separately.