人臉偵測常常作為人臉追蹤與辨識的前置作業，是一個複雜且困難的研究課題，其結果足以影響整個系統的效能。本論文提出一個以膚色為基礎之互補人臉偵測策略。首先在離線時統計人臉膚色的色彩資訊，並且使用一些形態學演算法加以描述。然後當系統從外接設備擷取到影像，便可以根據色彩資訊分離出可能有人臉存在的區域。接著使用兩種功能互補的策略來偵測人臉。第一種策略是新的人臉特徵抽取方法，利用膚色資訊來顯露出臉上的特徵圖樣，並且用一些幾何關係的限制加以過濾。第二種策略是使用可調式搜尋視窗框出可能的人臉候選，此視窗只需在固定解析度的影像上搜尋，可以自動的判斷移動方式並且調整視窗大小。最後是採用主分量分析來確認這兩種策略所找出來的人臉候選。 我們的實驗證明本論文所提出的方法確實可行且可靠，在少許的條件限制下，對於環境採光變化、不同大小之人臉、人臉各種姿勢與表情、人臉有部分被遮蔽、及複雜背景等問題都可以有效的處理。此外，本論文還可以偵測出多個互相交疊的人臉，這是從前以膚色來偵測人臉的研究所無法克服的。 Face detection is a complex and difficult problem. To serve as a prior step in face tracking and recognition, it is the most important process involved. In this thesis, two complementary strategies are presented to detect human faces in images based on skin color information. The proposed system consists of three major parts. The first part is to search for the potential face regions by off-line statistic information of skin color. The second part performs face detection by two strategies with complementary capabilities. The proposed first strategy can quickly extract facial features and choose some of them as the candidate of faces to be confirmed by some effective geometric constraints. In the second strategy, adaptive search window is utilized to locate face candidates. It can determine the moving way and adjust window's size automatically to adapt face region. Lastly, the algorithm of principal component analysis is adopted to verify the candidates of face obtained by the previous processes. Experimental results reveal the efficiency and feasibility of our proposed approach. Under fewer constrains it can conquer difficulties, such as different lighting conditions, sizes of faces, variable orientations, facial expressions, partial occlusion, and complicated background. Moreover, we can also handle the case of occlusion of multiple faces.
