基於主成份分析法與灰關聯分析法之動態人臉辨識

Abstract

人臉辨識系統廣泛地應用於身分認證、門禁管理與人機界面等領域，近年來由於「智慧生活」科技的提倡，人臉辨識技術已延伸至人與機器最佳化介面之應用。此外視訊會議、影像內容檢索與醫學影像處理等方面，亦是其重要之應用領域。 本篇論文分為人臉偵測和人臉辨識兩大部分。在人臉偵測的部份，我們利用膚色分割和連通成份的方法找出人臉候選區，再使用色彩分析的方法從人臉候選區中尋找眼睛和嘴唇的特徵，最後再使用眼睛和嘴唇的幾何條件關係去定位出正確的人臉位置。在人臉辨識部分，我們提出一套結合主成份分析法與灰關聯分析法的人臉辨識方法，此方法的架構分為以下三個階段：首先，在影像前處理的階段，我們使用二維小波轉換，對輸入影像做資料壓縮的處理，接著，利用主成份分析法將壓縮過的人臉影像，投影到低維度的子空間中，計算出具有代表性的特徵臉，最後，再使用灰關聯分析法，來辨識出正確的人臉圖片。 為了驗證本篇所提出的方法，在靜態辨識實驗中，我們使用ORL人臉資料庫，做了一些分析和比較的實驗，實驗結果證明，在40人條件下，訓練樣本為五張時，可以得到91.6%的辨識率。而本篇方法在動態辨識實驗中以不同距離拍攝人臉，在30人條件下，可以得到八成以上的辨識率。

Face recognition systems have been utilized in areas such as biometric identity authentification, acess surveillance, and human-computer interface. More recently, because of the promotion of “intelligent life”, the use of face recognition techniques has been extended to optimizing the human-computer interface. In addition, video conferencing, image content indexing and medical diagnostics are other applications for face recognition. This paper first discusses the face detection part, and then discusses the face recognition part. For the face detection part, we used skin color segmentation and connected component method to extract a face candidate. Then color analysis was used to identify the features (lips, eyes) of the face candidate. Finally, measurements related to the eyes and mouth were used to locate the position of the face. For the face recognition part, we present a hybrid face recognition method, which combines Principal Component Analysis and Grey Relational Analysis. The proposed method consists of three stages. First, during preprocessing, we performed a Discrete Wavelet Transformation for data compression. Second, using Principal Component Analysis to project the input images into a low dimension subspace, we calculated the representative eigenface. Finally, we used Grey Relational Analysis to recognize the face images. To confirm our proposed method, we performed static and dynamic recognition experiments for analysis and comparison. ORL face databases were used in the static recognition experiments. Our database contained 40 people, and for each person, we selected 5 training samples. Using these training samples, we obtained an accuracy rate of 91.6 percent. In dynamic recognition experiments, we were able to obtain greater than 80 percent accuracy for 30 people under different distances.