臺灣博碩士論文加值系統

　　在超低位元壓縮的領域中，模式壓縮（model-based coding）是一種有較佳壓縮效果的方法。這種方法被設計來對頭和肩膀的影像的序列的編碼，根據頭部的結構特徵可用大量預備知識來加以模塑。我們用一個三維的頭部模型來描述出這些影像，並將其傳送。在重建頭部模型時，臉部特徵點的萃取是個重要的步驟。
　　在本論文中，我們提出一個四階段萃取方法來找到61個特徵點，這四階段萃取方法為：一.訂出臉部器官的範圍、二.找出器官角落的點、三.找出器官外型、四.萃取特徵點。這61個特徵點分佈於頭髮、下巴、眉毛、眼睛、虹膜和嘴巴。我們使用幾個1秒30張的連續影像來做實驗，實驗結果我們的四階段萃取方法相當有效，近百分之九十的特徵點能被正確的找到。

　　The model-based coding is recognized as the most effective way of coding invery low bit rate. It is designated to encode head-and-shoulder image sequences, in which whose structural characteristics can be modeled based on priort knowledge. That is, a 3-dimensional head model should be sketched out for the head-and-shoulder image sequence to be transmitted. The facial feature points extraction is one of the key techniques in recapitulating the model.
　　In this thesis, we propose a four-phase extraction scheme: (1)facial-expression area locating, (2)corner point finding, (3)contour tracing, and (4)feature point extracting, to locate 61 reature points in one image. These points are distributed in six facial-expression objects of a face, they are: hair, chin, eyebrow, eye, iris and mouth. We captured several image sequences for experiment, each sequence has 30 images for one second. The experimental results show that our four-phase scheme is effective, near 90% feature points are located accurately.