基於景深之自動化透明度分析

Abstract

影像透明度分析(Image matting)指的是分析影像中每一個像素是屬於前景、背景、或是兩者的混合，而每一個像素由多少比例的前景和背景組成可用alpha這個參數來表示。過去各種影像透明度的方法均需要一些使用者輸入，例如，在前景和背景劃上筆劃或是用trimap來表示前景、背景和不確定的像素，才能得生出合理的alpha值。我們在這篇論文提出一種新的方法來自動算出每個像素的alpha值，且不需要任何使用者的輸入。 我們假設背景離前景有一定的距離，而且照片是在較淺或有限的景深下拍攝，因此前景清楚而背景模糊。我們首先利用這個假設產生出一個粗略的trimap，再用這個trimap來初步估計所有不確定像素的alpha值，最後再將整張照片的alpha值做最佳化，便能得出精確的alpha值。

Image matting is the problem of determining each pixel in an image whether it is foreground, background, or a mixture of foreground and background while 'alpha' value is the parameter of portion of foreground and background in each pixel. Previous matting approaches are constrained by requiring certain amount of user input, for example, foreground/background strokes or trimap, to generate reasonable results. In this paper, we present an approach to automatically generate an alpha matte without any user input. We assume that the background is distant from the oreground and images are taken in narrow or limited depth of eld so the foreground is in focus while the background is blurred. We rst generate a rough trimap under our assumption and then use this trimap to initially estimate alpha value of each unknown pixel. And after further optimization, a high quality alpha matte can be achieved.