人臉五官定位是人臉識別的先置作業，也是追踪人臉及表情辨識的重要步驟。會遇到的問題包括環境光線的變異、人臉方向的改變、複雜的背景，及相機本身的特性。過去有許多方法被提出來，依需求來解決上述的困難。我們的人臉及五官定位將應用在固定相機的門禁系統上；因此要解決的問題大致只有光源的變異而已。 本論文包含三個主要議題: (i) 人臉定位，(ii) 鍊碼追踪與符號描述， (iii) 眼鼻口定位。第一個主題將探討擷取人臉的方法。第二主題是將人臉區域的輪廓轉為鍊碼和符號描述，提供匹配眼睛所需的符號描述資料。第三個主題探討如何利用符號描述來找出雙眼，之後再利用五官模板來找到鼻口。 用膚色偵測，是最直接取得人臉定位資訊的方法，不受複雜背景的影響，也不需要經過學習與分類。所以我們採用膚色定位的方法，將偵測出來的膚色區塊交給下一步來判定人臉。判定人臉的方法，使用邊偵測後，鍊碼追踪再轉成符號描述群組並檢測群組是否存在人臉五官比例，這樣的做法能獲得較可靠的判定結果。 經過判定的人臉，就可以使用符號描述群組的索引找雙眼最後再以人臉模板匹配出鼻和口。我們定義偵測區塊要包含臉部器官的70% 範圍，且偵測區域面積不大於該器官面積的2 倍，才算偵測到該器官。總共使用56張照片，共有63個人臉做實驗，而五官定位方法經實驗測試在戶外偵測左右眼、鼻、口及取得各五官範圍可達71.4% 的偵測率，在室內的大頭照則可達到77.1% 的偵測率；採用MUCT人臉資料庫取用150樣本偵測眼、鼻和口符號群組達74.7% 的偵測率。未來將對頭髪和眼鏡結構來進一步探討，就可以再提高定位的機率。The face-feature location is a pre-task of face recognition; it’s also an important procedure for facial expression and face tracking. The encounted problems in the face-feature location include the change of ambient light, the direction the face, the complicated background, and the sensation of camera. Many methods have been proposed to resolve these problems. This paper consists of three main topics: (i) face detection, (ii) chain code tracking for faces, (iii) and face-feature location. The first topic discusses an approach of face detection. The second topic describes how to transfer the face contour and face-feature shapes into chain code representation and symbol description to match and verify eyes. The third topic utilizes the symbol description to find the locations of nose and mouth by face-feature template. The skin color is the most useful characteristic for detecting faces without affecting by the complicated background. We here adopt the skin color to extract face candidate following face verification. The verification procedure uses edge detection and chain code representation to describe the face. We feel that based on such an approach, we can get more reliable results. After the face extraction, the eyes are found by utilizing the symbol description and the nose and mouth are found by matching the face template. Experiments show that eye, nose, and mouth were correctly located at a rate of 71.4% in outdoors, and at a rate of 77.1% for indoors. In the future, we will explore the structure of hair and glasses to enhance the detection rate.
