Feature Dimensional Reduction For Fast Face Recognition

Among various identification based biology character technology, face recognition technology has been extensively applied to abroad range of field due to its unique superiority compared with others. As the key issues in face recognition, feature extraction and expression have received considerable academic attention in several years. On one hand, different kinds of face feature expression and dimensions have a direct effect on face recognition accuracy. Usually, in the same face expression method, the higher the dimension, the accurate the face recognition is. On the other hand, the dimension of feature extraction will direct affect the real-time of face recognition. High dimension will take longer time for face recognition and result in lower real-time.In this dissertation, with the aim to satisfy the need for realizing timely face recognition, we analyze the feasibility of Principle Component Analysis (PCA) in reducing the dimensions based on the describe of face feature expression and recognition of local descriptor. The mainly works and contributions of this dissertation are as follows:(1) The face feature expression and recognition based on PCA, LBP and Local Derivative Pattern (LDP) are investigated thoroughly. We analyze their characteristic and evaluate the performance of face recognition for the three methods. Compared with global descriptor based on PCA, the face feature expression based on local descriptor can describe face feature more effectively. Hence, the later can improve the rate of face recognition substantially. In comparison with one order local descriptor,the face feature expression with high order local descriptor has much more spatial information and thus can further improve the rate of face recognition.(2) As PCA has advantage in lowering dimension, we explore a new face recognition method which combines the characteristic of local descriptor and PCA. The face recognition speed is enhanced by using PCA to low the dimension of face feature extracted from local descriptor. We check our new method in FERET. The results demonstrate that under the same recognition rate, the proposed method can lower the feature dimensions in the stage of face recognition and reduce computational complexity in the process of face recognition.