Loading...

Face Recognition in Subspaces Based on Nonlinear Dimension Reduction

Mohseni Takallou, Hadis | 2013

564 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 44946 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Kasaei, Shohreh
  7. Abstract:
  8. In many applications in human society, there is a need for identity recognition of people.Different methods have been developed for this purpose while using the biometrics is one of the major interests. The biometrics measure the unique physiological, anatomical and behavioural characteristics of people. Among them, face is an interesting biometric which have important advantages over other biometrics and face recognition is known as the most common method that people utilize to recognize each other. However, face recognition suffers from factors such as changes in head pose, illumination and face expression which influence the efficiency of recognition methods. The core of many recent researches focus on strengthening face recognition methods to variation in mentioned factors.In this research, the aim is to propose a face recognition method which is robust to head rotation. The basic idea is to use dimension reduction methods that map high-dimensional face images to the low dimensional space where variation in face can be described with few parameters. Based on the evidences that imply the existence of a manifold for face images, we have used the multi-linear and non-linear dimension reduction methods for the face recognition application. This have been done through three proposed algorithms when every algorithm complete the previous one. In all these algorithms, the aim is to separate the identity and pose factors of a face image. The first algorithm emphasises on using multi-linear decomposition for extracting an identity-independent pose manifold based on a pose-based dissimilarity measure. In the second algorithm, approximation tool has been combined with multi-linear decomposition to make the mentioned separation more precise and to find the mapping between the image space and the pose manifold. Making the assumption that there are face images in several poses, the pose-independent mappings are computed that are used for both pose estimation and face recognition aims. In the third algorithm, the approximation tool is used to find the mappings between the rotated face images and the corresponding frontal face images. These mappings are identity-independent and bring up the possibility to recognize people based on just one frontal image. The recognition rate higher than 90% on three FacePix, Oriental and CMU-PIE databases in different experimental conditions shows the efficiency of the proposed algorithms. Although the recognition rate in 3D models is higher than the proposed algorithms because of using 3D information, the recognition rate is comparable or higher than similar 2D image based methods
  9. Keywords:
  10. Dimension Reduction ; Face Recognition ; Human Pose Estimation ; Poses Manifold ; Multi Linear Dimension Reduction ; Nonlinear Dimension Reduction

 Digital Object List

 Bookmark

...see more