臉部特徵變化之自動偵測

Abstract

這一篇論文提出了一個新的方法, 可以自動地在任意表情的臉部影 像中擷取出可變動之特徵. 所謂可變動的特徵包括眼睛和嘴巴張開的大 小. 這一個方法的應用範圍相當的廣泛, 如人機介面, 電腦視覺等等. 整 個方法分為三個階段, 第一個階段是先在一張" 無表情" 的臉部影像中找 出特徵的Bounding box, 並擷取出特徵. 第二個階段則在" 任意表情" 的 臉部影像中, 找出特徵的Bounding box, 並擷取出特徵. 在這二個階段中 尋找特徵的Bounding box所用的過程是不相同的. 主要的原因是在無表情 和任意表情的臉部影像中特徵的形狀變化是隨意的. 而第三個階段則分析 這些所擷取出來之特徵, 並轉成數量化的描述. 由實驗結果得知, 在適當 的光線, 和臉部沒有太大的傾斜時, 這一個方法可以得到良好的結果. This thesis proposed a new approach to automatic extract variable features, such as eyes and mouth, in a facial image of arbitrary expression. This approach has many applications, such as man-machine interface, computer vision,etc. The overall approach consists of three separated but interrelated stages. The purpose of first stage is to locate the bounding boxes of the features and to extract the features in an expressionless image. In the second stage, similar procedures are carried out for an arbitrary expression image. The two stages are not the same, because the changes in the features' shapes are unknown. Finally, the approach generates quantitative descriptions of extractedfeatures. According to the experimental results, when the lighting condition is normal and the tilt of face is small, the approach can produce reasonably goodresults.