Research On Expression-robust Of 3D Face Recognition

Currently, with the rapid development of global informatizion, effective identification and accurate certification of identity have been widely concerned. As an important branch in the field of biological feature recognition, face recognition has become the focus of the domestic and foreign research scholars, because of its unique advantages and highly academic value. After decades of development, under some constraint conditions, a very ideal recognition rate is achieved by 2D face recognition technology based on apparent features of images. However illumination, pose and other factors are the biggest obstacles to the forward development of 2D face recognition technology.An opportunity for researchers is provided by the rapid development of 3D data acquisition and storage devices, thus 3D face recognition technology is expected to solve the difficulties presenced in 2D face recognition. After in-depth analysis of 3D face data and closely focus on pose and expression problems, a 3D face recognition algorithm is proposed by this paper, which has good robustness of expression change. The main research contents of this paper are as follows:For scanning error of 3D data, such as disorder, spikes, holes, an effective pre-processing method is proposed. First, local search method is used to locate the nose tip point, and cut the face with a certain radius, second, principal component analysis(PCA) is used for pose correction, third, multilevel B-spline interpolation algorithm is used for surface fitting, finally weak noise and reduce the within-class difference by smoothing.To make up the defect of using single and insufficient feature information, a variety of features and information are comprehensive used in this paper, to overcome the influence on 3D face recognition which expression change brings, a algorithm is proposed by fusing the depth image and three rigid areas. First get the depth image, then expression changing threshold is obtained by extracting and training the image of mouth region. If expression change value is greater than the threshold, the part above the mouth is retained, else the whole image is used, next match the depth image using 2D-PCA algorithm.Three regions contain whole nose, left cheek, forehead, are choosen to match with each other using ICP algorithm baesd on k-dimension tree acceleration, finally both the depth image’s similarity and three rigid regions’ similarity are fused in decision level.Through the judgment of depth image of mouth, the part above the mouth is retained if expression change value is greater than the threshold, so the non-rigid deformation caused by expression is weaken. Nose area is extracted accurately to increase the inter-class difference, and the information features contains is increased by adopting three rigid areas. Test results on database show that our algorithm has better robustness against expression change than other algorithms.