2d And 3d Face Recognition In A Number Of Key Issues

Face recognition becomes more and more popular in theses years for its convenience, straightforwardness and effectiveness. While with the development of face detection, recognition and modeling, the traditional two-dimensional face analysis are confronting challenges including Pose, Illumination and Expression (PIE) problems, efficient feature extraction and curse of dimensionality. Since 3D human faces provide the information which was lost in the 2D images; better describe the facial features; successfully achieve rigid invariance, it has received more and more attraction from researchers. However, 3D face methods also have the following problems: inconvenient acquisition of 3D faces, data representation and alignment, and less efficient 3D face recognition methods etc. Focused on these problems, some creative research and works have been done in this thesis.1) Aimed at the problems of viewpoint and pose in 2D face recognition, we propose a linear discriminative method based on label constrained graph partition. We divide each class into subclasses using the distances between sample pairs and their label information. In addition, Most Discriminant Subclass Distribution (MDSD) Criterion is proposed in this paper to select the optimal partition of the subclasses.2) We propose a method to localize 3D facial feature point based on Bezier Surface model for 3D face representation and recognition. Local geometric structure around feature point is represented by Bezier Surface and used for discriminating different feature points. Based on the different complexities of local structure we employ different interpolation accuracies to achieve the tradeoff of computational cost and localization accuracy. Follow the scheme of "local search plus global constrain", 60 feature points on 3D point cloud can be located accurately.3) The new 3D face recognition method is based on Facial Structural Angle from 19 important feature points and Local Region Map from the estimation of local structure around feature points. The experimental results show that it is invariant to Pose and Illumination and robust to facial Expression.4) We propose a method of synthesizing 3D face via frontal and side view photos. The generic 3D deformable model is projected onto the frontal and side view plane, hence the differences of feature points positions between projected image and real image are obtained. Then we use this information to calculate the actual movement of feature points from generic face to specific face. The free points on the 3D face mesh are defined by Linear Movement Criterion of Free Points.During master study, I also finished a Railroad Detection System in which we proposed a Local Line Pattern Description to detect the line in railroad images.