Face Recognition Based On 3d Features

Automatic face recognition (AFR) has been studied for over 50 years. It has become one of the most active research areas in computer vision and pattern recognition. Much progress has been made on its theories and algorithms in the past years. But most of them are based on 2D face images. It has been demonstrated that the pose and illumination variances of face images influence seriously the recognition performance, and the images of the same individual with different poses or illuminations have less similarity than the images from different individuals but with the same pose or illumination. Therefore, 2D face images are much sensitive to face pose, illumination and expression, which is one of the greatest remaining research challenges in face recognition.With the development of computer hardware and 3D information acquisition technology, it is likely and convenient to acquire the 3D data of an object. Compared with the 2D face image, the 3D face data contains more spatial information, which is inherent property of the face and is robust to the uncontrollable environment where 2D appearance can be affected. In recent years, the 3D face recognition has attracted more and more attention, and has become one of the valuable research topics. Meanwhile, many 3D face recognition algorithms have been presented. In this paper, we summarize and study the state-of-the-art 3D face recognition algorithms and the 3D face models, and focus on discussing the face recognition technologies based on 3D features. We mainly address the problem of the robust 3D face representation and recognition, and also propose some resolvents for 2D face recognition under some certain circumstances. The main work of this thesis includes:1. 3D face recognition based on geometrical features3D faces are informative, and contain many kinds of face features. How to design and extract the robust and effective features for face representation is an all-important issue for face recognition. Based on the research of 3D face data acquirement and normalization, we discuss the representation and similarity metric measurement of 3D face geometry features, extract five kinds of face features, and analyse their representative and discriminant ability by experiments. Different features provide different discriminative information and also own different discriminative capability, while a single feature only provides the limited discriminative information and local applicability. So we employ a linear weighted strategy based on Fisher linear discriminant analysis to fuse multi-features for recognition, which promotes the recognition performance to some extent.2. 3D face recognition based on Local Binary Pattern (LBP)3D faces are robust to facial poses and illuminations, while sensitive to expressions, and facial expression can distort the 3D mesh largely. In order to weaken the effects of facial expression, and extract the expression-robust 3D face features, we study the 3D face recognition algorithm based on LBP, and explore the 2D LBP operator to 3D form. We discuss the definition of 3D LBP operator, the division strategy of 3D facial surface, and the matching algorithm based on the statistical histogram for 3D face recognition in detail. Experimental results demonstrate that it is feasible to apply the LBP framework on 3D face recognition, and the 3D face recognition algorithm based on LBP is robust to facial expressions to a certain extent.3. 3D face recognition based on sparse representationThere are many kinds of 3D face features, and how to extract the robust and effective features for face representation is a crucial problem. In this thesis, we make more efforts on sorting and selecting 3D face low-dimensional geometrical features, and present the 3D face recognition algorithm based on sparse representation together with the sparsity evaluation for facial signal. Experiments demonstrate that our method can describe the 3D face more exactitude, and improve the robustness of recognition algorithm for face expressions.4. Face recognition based on HaarLBP assisted with 3D virtual featuresConsidering 2D face images are sensitive to face poses, illuminations and expressions, while the 3D faces are a litter robust to these varieties, we propose a face recognition algorithm fusing 2D HaarLBP and 3D virtual geometrical features. First, in order to get more detail information from 2D face images, we study and propose a HaarLBP based 2D face representation and metric measurement. Considering the limited representative ability of 2D images, we estimate the 3D shape according to the given face image based on 3D morphable model, and then extract robust 3D face features as complement for 2D features. At last, we fuse the 2D and 3D face features to address the recognition task, and conclude that our method is a litter robust to face poses, illuminations and expressions.5. Pose-invariant face recognition based on a single viewThe face recognition algorithms based on a single view are often subject to the lack of training samples, and so can't extract adequate discriminative information for classification. According to this problem, we reconstruct the 3D face according to the given 2D face image based on 3D morphable model, and then many virtual multi-pose face images can be acquired by 3D faces rotation and projection. Based on this, we can construct the face pose subspace for each individual, and extract the pose- invariant common features for face representation and recognition. This method can enlarge the face multi-pose samples, which guarantees the performance of multi-pose face recognition algorithms, and the experiments demonstrate the validity and robustness of our method to face poses.