3D Face Recognition By Geometric Feature Fusion

Face recognition technology has become one of the most popular topics in the field of biometric recognition due to its friendly, convenience, concealment. After almost five decades of development, the face recognition based on2D image has made lots of achievements and become more and more mature. However, the2D face recognition can’t solve the problems of posture and light variation because of the data format of2D image.3D face recognition is supposed to provide an opportunity for overcoming the problems2D face recognition faced with. Our group has done lots of work about3d face data acquiring based on binocular stereo vision technology and extracted geometric features from face for recognition, which has made some progress.However, the3D face data which acquires based on binocular stereo vision technology includes noise. To make the3D face model more smooth for improving the final recognition rate, this paper focuses on the point cloud filter and more geometric features extraction and fusion. The main work and contributions are as follows:(1) Acquisition of3D face data and point cloud filter. To deal with the noise problem existed in the acquiring facial data, we use the point cloud bilateral filtering technology to filter for the XMU-3DFED.The experimental result show that the bilateral filtering algorithm can not only perform well on the small-scale noise, but also keep intrinsic characters of the face surface.(2) Preprocessing on3D face model. To solve the problems of useless data, posture variation, format conversion requirement, we respectively provide the procedure of face segmentation, pose correction and surface reconstruction. What’s more, we also study the3D point cloud model matching ICP algorithm, propose two methods:KD-tree and Delaunay triangulation to improve the efficiency of ICP algorithm.(3) New geometric feature extraction for face recognition. This paper mainly extract two types of geometric feature, the first type is the13-dimensional geometric feature G that is composed of curvature, volume, distance, angle extracted from the region of the nose; the second type is made up of three kinds of features:distance feature D, area feature S, angle feature A, which are extracted from the triangular consisting of the nose-tip and two points form iso-geodesic contour. The experimental result proves that the new geometric feature can distinguish the faces.(4) Proposed the experiment system of geometric feature fusion. We adopt the decision level fusion for two types geometric feature and three kinds of features from the second type. The recognition rates of the decision level fusion were verified on XMU-3DFED and ZJU-3DFED. The experiments show that compared to the single geometric feature method, the decision level fusion method get higher recognition rate, in which the fusion of geometric feature G and the distance feature D get highest recognition rate on ZJU-3DFED which get5percentage points more than the best single geometric feature on rankl and achieve97.14%on rank3.