| 英文摘要 | The objective of face recognition is to enable computer to determine a person's identity according to his face data, such as image, 3D shape, video and etc. Firstly, the most essential and discriminant features are extracted from face data. Then the correspondence between face data and identity can be established by some classifiers. With the development of face recognition technologies, Many efficient methods have been proposed, with nearly 100 percent recognition rate on some standard databases. However, there still exist a lot of challenges in practical applications, which mainly caused by illumination, pose, expression, age and etc. Nowadays, while mass researches are focusing on llumination and pose problems, face data acquisition is almost neglected, which is one of the most important factors affecting the quality of face data. In some applications, face data may be acquired by different imaging devices, e.g. enrollment by visual (VIS) images and authentication by near infrared (NIR) images to resist various illuminations on-site. Due to the different imaging mechanisms, even under the same environment,the face data collected by various devices are different, which will decrease the performance and restrict the application of face recognition. Face data collected by different devices or collected under various illumination, pose and etc. are all called "heterogeneous face data", which represents a unique identity while being of different quality. This thesis mainly focuses on NIR-VIS heterogeneous face recognition problem. And accordingly, some effective methods are proposed for heterogeneous face recognition. This study involves a lot of key problems in computer vision and face recognition, e.g. face pre-processing, template matching, subspace analysis and multi-variates regression. The main contributions of this thesis are shown as follow: 1 We present a new problem, that of face recognition between NIR and VIS images. 2 We propose a "Canonical Correlation Analysis" (CCA) based regression framework to solve the problem of heterogeneous face matching. NIR, VIS and 3D face matching are taken as examples to illustrate the effectiveness of the method. 3 We propose a local feature matching algorithm to solve the problem of generality in conventional heterogeneous face recognition methods. The algorithm can handle the partial occlusion of face images and achieve good performance on MBGC portal challenge. 4 We build a "Enhanced Near infrared" (ENIR)... |
修改评论