Research On3D Face Recognition Based On Invariant Features

Face recognition is a key subject of the research on biometrics. Due to the strong academic background and broad application prospects, it has become the focus of domestic and international top research institutions and scholars. With the rapid development of the huge number of related disciplines, such as image processing, pattern recognition, computer vision, statistical learning, cognitive science and psychology, face recognition has been in great need and necessary of national security and public safety, and becomes one of the hottest issues in that field. However, face recognition based on2D face images is still challenged by the large change of illumination, pose and expression after received more than10year’s research and its recognition rate is still far away from satisfaction under the change of the above three factors.With recent progress in3D sensors, more and more organizations have committed to using the three-dimensional data-aided face recognition and made a good result.3D face recognition has potential to overcome the difficulties of the image-based face recognition caused by the variations of illumination, facial posture and expression etc. Effective feature extraction of3D facial surface is one of the core technologies for3D face recognition. To extract the most distinctive3D facial features of the same individual, which is different from the general public, is the first and crucial phase for a highly efficient face recognition algorithm. In this dissertation, based on the detailed and in-depth analysis,3D facial feature descriptors for3D face recognition has been studied and discussed. Combined with the surface shape characteristics of3D facial images and statistical learning theory, we can extract the discriminant features with invariance. The main research content and innovative work are as follows:1. Proposed an automatic pre-process technique and regional segmentation approachBased on the correspondence between the2D texture channel and3D data, we first roughly extract the facial area. The facial central stripe is used to detect the nose tip and remove some clutters. Pose correction and registration based on Axis-angle representation can be fixed and the facial pose is estimated and put into a canonical framework. Shape index and facial geometrical constraint are introduced to segment the main organ regions. The proposed3D face automatic pre-process method can contribute to the realization of the subsequent feature extraction and classification algorithms. It can not only improve the computational efficiency, but also improve the quality of the3D face data in the original input.2. Proposed three kinds of3D facial feature descriptors1) Bending Invariant is proposed for describing the3D facial information. The transformations of expression on a facial surface can be considered as non-rigid transformations, and empirical observations show that facial expressions can be modeled as isometric transformations. Bending Invariant can be used to construct a signature for facial surfaces. This descriptor can transform the non-rigid deformation caused by expression into the rigid transformation by performing an ISOMAP on a reduced set of points and interpolating on the full set of points. Experimental results show that Bending Invariant can effectively improve the recognition performance of3D face data with large changes in the expression and overcome facial existence of the surface deformation caused by the facial expressions.2) Bounding sphere representation (BSR) is proposed as3D facial descriptor for3D face recognition with pose variations. In order to better reflect the facial surface shape and increase the discrimination, BSR is used for the aligned3D point cloud data and has been justified as effective on3D information description. The descriptor is imaged as the projection of the relative position of a facial point cloud into bounding spheres centered as the centroid point of the face. The BSR can preserve the lower descriptiveness, and reflect the facial characteristics on a rotational spherical domain intuitively compared with the other descriptors, which makes it insensitive to large expression and pose variations. Experimental results show that BSR can retain more geometrical information, extract the descriptor with posture-related characteristics, and overcome the influence of self-occlusion.3)3D scale invariant feature transform (3D SIFT), with affine, rotation and scale invariance, is proposed for describing3D facial surface, which can effectively detect key points based on the3D facial surface information and encode the facial surface information to describe the key points. The intuition to add depth is that the changing of the depth value of the key points may give us more information to discriminate the subjects, which are visually similar to each other but are different because they are in different level of depth. Experimental results show that3D SIFT descriptor can effectively characterize the distorted and deformed images, restore the essential characteristics of the3D facial surface, extract the discriminant information and improve the recognition result.3. Proposed a robust group sparse regression model for discriminant feature extractionExpression, poses, occlusions and corruptions are common problems that significantly influence the accuracy gain of the3D system. We introduce the theory of low rank and group sparse representation to analysis our feature representation. Combined with nonlinear corruption constraint and a supervised Spectral Regression constraint, a lower intrinsic dimensional feature vector can be extracted based on a novel optimization model. Hair occlusion and data corruptions can be patched handled simultaneously. The extracted low-dimensional features are more discriminative, robust and generative. Experimental results show that3D facial intrinsic feature vector based on RGSRM has a good dimensional data representation capability and show superior performance.4. Designed and implemented a3D face recognition systemFor the purpose of3D face data acquisition and proving the generality of these algorithms proposed upon, the3D face recognition system is based on the stereo vision and structured light. The software function modules are composed of image capturing, camera calibration,3D face model generation,3D face model smoothing,3D face model pre-processing,3D face description and feature extraction. All work lays a good foundation of further research and development of3D face processing and recognition.