影像透明度分析之三元圖生成法

Abstract

影像透明度分析（Image Matting）是這幾年來影像處理的熱門題目之一，幾乎每年都有相關的論文發表。在大多數的方法中，三元圖（Trimap）都作為必要的輸入之一，但是三元圖的取得方式在文獻上卻沒有太多著墨。然而，三元圖的準確性會大大的影響透明度分析的結果，所以如何針對該影像得到一張好的三元圖是相當重要的。因此本篇論文提出一個利用影像分割（Image Segmentation）的結果，對影像內容作分析而自動產生三元圖的方法。 首先，針對影像分割後物件的邊緣做擴增（Dilation）形成三元圖初步的猜測，接著再對這張三元圖做透明度分析即可得到影像大略的結構。透過對這些結構的分析，動態調整筆刷的寬度，即可產生一張針對影像內容變化的三元圖。最後再針對影像中細小的結構做補強來得到最後的三元圖。實驗結果顯示我們提出的方法，可以有效地加強影像透明度分析的結果而由於準確的三元圖減少了需要運算的區域，也大大壓縮了影像透明度分析所需的時間。

Digital image matting is one of the popular topics in image processing these years; there are almost two to three theses published every year. In most of these methods, they use trimap to be one of the inputs, but don’t concern about how to get the trimap. In fact, the accuracy of trimap will much affect the result of image matting, so generate a good trimap for an image is very important. In this thesis, we present an automatic content aware trimap generation method based on the result of image segmentation. First, we dilate the contour of the result after image segmentation to be an initial guess of the trimap. We use this trimap to perform image matting; the result of matting will give us a rough structure of this image. By analyzing this information, we can use the method called dynamic brush width to construct a content aware trimap. The final trimap will be generated by this trimap with small structure enhancement. The experimental results shows that the trimap generated by the proposed method can effectively improve the result after matting, and because the trimap accurate reduce the region needing to be process, the matting is also accelerated.