Face Recognition Using Principal Facial-Factor Analysis

• Main Authors: Hsun-Li Chang, 張循鋰
• Other Authors: Tai-Wen Yue
• Format: Others
• Language: en_US
• Published: 2002
• Online Access: http://ndltd.ncl.edu.tw/handle/95625543293054287479

碩士 === 大同大學 === 資訊工程研究所 === 90 === Owing to the urgent demand of the security system is increasing in recent years, the techniques of personal identification become more and more important. Recently, the growing research interest of personal identification is focused on the biometrics information such as retina, fingerprint, voiceprint and human face. The biometric of an individual are exclusively and uniquely. Many methods have been proposed and some commercial products have been presented to the public. However, they are often restricted to particular conditions and the mode of operations.Therefore, the most convenient and feasible way to personal identification is by human faces. In machine face recognition, how to represent a human face effectiveness and efficiency is a main topic for research. Many approaches have been proposed. The most popular template-based approach is eigenface method. The main idea of this approach is to consider a face image is a linear combination of a set of basis face images. It applies PCA (principal component analysis) to find the set of basis images. However, the eigenface approach may be too rough to describe the detail on a human face. We consider that human faces are composed of some distinguishable facial factors, such as the style of hair, eyes, nose, and mouth. Therefore, instead of performing face recognition based on the eigenfaces retrieving from whole faces, we gave a solution that was based on eigen-facial-factors. The method to retrieve the eigen-facial-factors from a set of face images is called principal facial-factor analysis (PFFA). In short, our approach provides PCA on segmented blocks of face images to extract the main components of facial factors that constitute the different faces. Using PFFA to represent a face image, the dimensionality of the original data will be greatly reduced, and the resulting feature vector will be more beneficial for facial recognition. In the thesis, the concepts of PFFA will be detailedly discussed. A comparison with classical eigenface method will also be given. The experimental results revealed that the recognition error rate of the proposed method is superior than the classical eigenface approach.