本論文中主要討論變形演算法在人臉影像上的應用,一開始選定兩張影像,其中一張待變形影像也就是欲變形的影像,另一張為一般的2D人臉影像。根據MPEG4表示虛擬人臉和動畫的臉部定義參數,分別找尋特徵區塊及特徵點,運用質心計算法及位移量定位出新的特徵點,最後使用狄勞尼三角化的特性完成變形程序。論文中我們限制參考影像的人臉所占整張影像的大小必須要大於待變形影像人臉 且參考影像儘可能五官特徵是清晰的。與過去研究人臉變形的方法不同,論文中是對兩張人臉中相對應的五官做變形,將人臉特徵點經過二維座標轉換及三角化後,對欲改變的五官做適當的調整。實驗結果可顯示所提出方法比過去使用貼圖合成的效果更好。
In this thesis, a method of human face deformation using image processing techniques is presented. First, two images are taken: the first is taken as the deformed image, and the second is taken as the reference image. According to the defined FDP (Facial Definition Parameter) to represent and animate the synthetic human faces in MPEG4, we first find out feature points from different regions, and next calculate the centers of mass and offsets to find new feature points. Finally, we use the property of Delaunay Triangulation to complete the deformation. In the thesis, the face size of the reference image is assumed to be larger than that of the deformed one. The proposed method is different from the past studies of the human face deformation. Our method is to deform the human face based on the reference image. After processing 2D transformation and Delaunay Triangulation for feature points, we make adjustments to the targeted facial features. The experimental result of the proposed method shows more effective than using composite map.