基於動態演化的數位影像前景不透明度分析法

Digital Matting Approach Based on Evolutionary Dynamics

吳秉竑(Bing-Hong Wu)

指導教授 ： 張隆紋

國立清華大學/電機資訊學院/資訊工程學系/碩士(2014年)

摘要

數位影像前景不透明分析(digital matting)是一個分析影像或影片中正確前景的問題。在這樣的問題當中，一張影像被視為不同透明程度的前景物件與背景的線性組合，數位影像前景不透明分析的目的是再一次將前景從來源當中分離出來。數位影像前景不透明分析是很多影像處理與製片應用當中的重要技術，其中最常見且廣為人知的技術便是利用背景屏幕的合成，也常常被稱做藍屏/綠屏合成。
在我們的方法當中，我們假設所有未知的前景都是已知前景的線性組合，同樣的假設也被運用在背景上面。基於這樣的想法，我們設計任兩組前背景組合的親合度，並利用演化動態的賽局來決定對於一個前背景未知的點來說，哪一組已知的前背景組合是比較重要的，讓這些前背景組合持續地演化到一個穩定的狀態以得到一個數學性的合理結果，最後藉由比較我們自己所提出的做法與另外4種常見方法來顯示出我們這個做法對於前景不透明分析的可行性。

關鍵字

數位影像前景不透明度分析 ； 動態演化 ； 演化賽局

並列摘要

Digital matting, or pulling a matte, refers to the problem of accurate foreground estimation in images and videos. In such problem, an image or a frame of video is considered as linear interpolation of foreground object and background with different opacity. The purpose of digital matting is to separate them from the source once again. It is one of the key techniques in applications for image editing and film production. One famous and commonly used technique of these applications is chroma key compositing, or blue/green screen compositing.
In our approach, we assume those unknown foreground and background intensities are linear combination of known foreground and background pixels accordingly. Based on this conception, we design the affinity value for any two of foreground/background color combination and use an evolutionary dynamic game to evaluate which combination is more significant to an unknown pixel. We can get reasonable results finally by letting these foreground/background color combinations evolve to a stable state. We compare the proposed method and 4 conventional methods to show its feasibility to solve matting problem.

並列關鍵字

digital matting ； evolutionary dynamics ； evolutionary game theory

參考文獻

[1] Martin A. Nowak. Evolutionary Dynamics: Exploring the Equations of Life. pp. 1-70, July 2003.

[6] Jue Wang and Michael F. Cohen, Image and Video Matting: A Survey, Journal Foundations and Trends in Computer Graphics and Vision archive, Volume 3 Issue 2, Pages 97-175, January 2007.

[8] Zhigang Xiang. Color image quantization by minimizing the maximum intercluster distance. ACM Transactions on Graphics (TOG) TOG Homepage archive, Volume 16 Issue 3, Pages 260-276, July 1997.

[11] Andrew N. Stein, Thomas S. Stepleton and Martial Hebert. Towards Unsupervised Whole-Object Segmentation: Combining Automated Matting with Boundary Detection. CVPR 2008, IEEE Conference on, June 2008.

[12] J. S. Y. Li and H. Shum. Video object cut and paste. ACM SIGGRAPH 2005, pp. 595–600