Research On 3D Face Recognition Across Expression

Automatic face recognition has extensive perspective of application in the fields of Homeland Security, Military Security, Public Security and family entertainment etc. Although human could recognize the human face and expression without striking a blowing, it is a big challenge for computer. The earliest research on face recognition can be traced back to the nineteenth century and since the technique of EigenFace was proposed in the nineties of the twentieth century, the vast majority of face recognition research have focused on the use of two-dimensional intensity images. However, face recognition based on face images is still challenged by the change of illumination, pose and expression after received more than 10 year's research and its recognition rate is still far away from satisfaction under the change of the above three factors.Recent progress in 3D sensors such as laser range finders has made acquisition of 3D human face data much more easily and cheaply which makes it possible to perform recognition based on 3D face data. 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. Based on the assumption that 3D face data have been captured, the illuminations have trivial effect to 3D face recognition. Due to its explicit representation of facial surface, the 3D face data has more information to conquer the change of pose than 2D images. However, expression is still a big problem. This thesis addresses itself to the task of 3D face recognition across expression and presents an efficient solution and a framework of 3D face recognition. The main contributions of the work are as follows: 1. Provided a thorough survey of the-state-of-the-art in 3D face recognitionFirstly, the background, conception and basic procedure of the 3D face recognition are given, and the 3D face data acquisition and representation are also concerned. Then, 3D face recognition approaches, categorized into three main groups: spatial matching methods, local feature based methods, and global feature based methods, are reviewed respectively. Also, face recognition using bi-modal of 2D+3Dis introduced briefly. Finally, we summarize the advantages and discuss the current challenges.2. Proposed an automatic pre-process technique and a posture localization approachBased on the analysis of surface curvature, we propose an automatic cropping method of facial region. Since 3D face data always contain some clutters, the facial region cropping could help to efficiently implement the subsequent steps of 3D face recognition. The facial mesh simplification introduced could improve the computational performance and the mesh smoothing remove the noise of the mesh. Furthermore, by the extraction of two feature points and the facial symmetry plane, the six degrees of freedom of facial surface can be fixed and the facial pose is estimated and put into a canonical framework.3. Proposed a Guidance-based constraints deformation (GCD) model to reduce the effect of expressionThe GCD model is to reduce the distortion caused by expression. Based on Poisson-Equation and a guidance face model, the GCD model could deform the non-neutral model toward neutral model which makes the intra-class models more similar. At the same time, a rigid-constraint is merged into the GCD linear system which retains the discrimination of inter-class models. Thus, the GCD model could improve the performance of 3D face recognition in presence of expression. By taking a Cholesky decomposition of matrix in Poisson equation, the solution could be obtained from back substitution which makes the computational cost reasonably acceptable.4. Proposed a representation scheme for 3D face -- Surface UnfoldingFor its irregular sampling in 3D space, 3D face surface cannot be processed by many orthogonal and subspace analysis toolkits directly, i.e. DFT, DCT, PCA, LCA. Surface Unfolding projects the 3D face to a 2D plane while retaining its intrinsic geometry properties of angle and area which issues a surface representation — Unfolded Image. The Unfolded Image is easy to process and the experimentsdemonstrate the unfolded depth image conveys richer discriminative information thanthe depth image.5. Proposed a facial surface matching approach — Partial-ICPThe Partial-ICP improves the ICP technique on three aspects which makes it more effective in surface matching. The initial point set is well selected to help two matching models have corresponding areas. The correspondence finding depends on the least distance from initial points to the triangle surface in the other mesh which accelerates the convergence of the iterative. The parameter, p-rate, is introduced to dynamically extract the rigid areas of facial surface at the same time of matching according to the shape distortion.