3D And Multi-modal Face Recognition

2D face recognition system can achieve good performances under controlled conditions. But its performance will drop drastically under the influence of some factors, such as illumination, pose, and expression variations. 3D face recognition can overcome or alleviate the influence of these factors. Multi-modal face recognition combined 2D with 3D information can be expected to obtain better performance. Several algorithms on 3D and multi-modal face recognition have been investigated in this dissertation.Firstly a 3D face recognition method based on iterative corresponding point (ICP) is presented. A clustering algorithm is proposed to eliminate point outliers from the facial point cloud. Then, the region of interest of the facial point cloud is extracted and transformed to pose coordinate system for coarse alignment. An approach based on symmetry property of facial surface is used to fill the holes of the facial data so as to improve the quality of facial data. And ICP algorithm is employed for fine registration. Finally, nearest neighbor classifier is adopted as the evaluation method. Experimental results demonstrate that the proposed algorithm have the capability of handling facial pose variation to some extent. The performance is still fairly good even when the facial data are of poor quality.A method of 3D face recognition which combines 3D local binary pattern (3DLBP) descriptor with generalized discriminant analysis (GDA) is proposed. Firstly a facial depth image is divided into regions. 3DLBP is used to extract histograms from these regions. All regional 3DLBP histograms are concatenated to a vector which is used as the feature of the facial depth image. GDA with modified Gaussian kernel is adopted as the classifier. Experimental results show that the recognition rate of 3DLBP combining with GDA is better than PCA and 3DLBP.Different fusion methods are used to combine facial depth images and grayscale images. LBP and local Gabor binary pattern (LGBP) are compared in detail. Experimental results illustrate that the combination of LBP and Fisher discrimiant analysis (FDA) is better than other methods. The performance after fusing facial depth image and grayscale images is better than that of unimodal ones. Methods based on LGBP cost more computation time and storage space, but have no advantages in performance compared with LBP based ones. A method which combined LBP descriptor with chain AdaBoost is presented for multi-modal face recognition. Thousands of regional LBP histograms (RLBPH) are generated from facial depth images and grayscale images respectively. Chain AdaBoost is utilized to select most informative RLBPHs. The selected RLBPHs are concatenated to a whole histogram. Then the corresponding linear subspaces are constructed by FDA respectively. Several methods are used to fuse 2D and 3D information. The experimental results demonstrate that very few RLBPHs selected by chain AdaBoost achieve fairly good performance. The performance will be improved further as the number of features increases.