使用邊緣梯度及水平集方法之臉部特徵擷取

Abstract

在此篇論文中，我們提出了一套臉部特徵擷取的系統，此系統流 程分成三個階段，分別為人臉區域偵測、臉部特徵區域偵測、臉部特 徵輪廓的擷取。第一階段人臉區域偵測，我們使用積分影像(Integral Image) 快速計算矩形特徵(Rectangle Feature)，並使用AdaBoost 演算法挑選出具有鑑別度的特徵構成分類器，透過Cascade 結構組合不 同效果的分類器以提升偵測效果。第二階段臉部特徵區域的偵測，我 們將重點放在眼睛與嘴巴區域的判定，透過眼睛周遭邊緣的密集程度 及距離資訊構成特徵，並使用支援向量機(Support Vector Machines) 建構出區域的預測模型；嘴巴的部分，由於膚色與唇色上的差異，透 過色彩分析可定位出嘴巴的區域。第三階段輪廓擷取，我們採用水平 集方法(Level Set Method) 來擷取特徵。在成功地偵測到人臉區域的 情況下，我們提出的臉部特徵擷取系統可以穩定地偵測到眼睛及嘴唇 的輪廓。

In this thesis, we propose a facial features extraction system which consists of three parts: face detection, facial features detection and features contours extraction. In the face detection, we use the integral image to efficiently calculate rectangle feature. The classifier can be constructed by the AdaBoost algorithm though selecting rectangle features with higher detection rate. We use the cascade structure, which is composed of the selected classifiers, to increase the detection rate. In the facial features detection, we focus on the regions of eyes and lip. The preliminary region of eyes is located by using the support vector machines (SVMs) with the information of gradient edge around the eyes. Then the preliminary region of lip is located by the color analysis due to the color difference between skin and lip. In the features contours extraction, we apply the level set method to extract facial features contours on the preliminary regions. Our method can stably detect lip and eyes contours under successful human face detection.