Research Of The Local Descriptors With Expression Variations For 3D Face Recognition

Face recognition technology has extensive applications in many social fields because of its intuitive,friendliness and high cost performance.The 2D face recognition technology has achieved a great success,but it is still difficult to solve the problems caused by illumination and posture.The 3D face data is less affected by illumination and posture,but it also faces the challenge of expression variations.In this paper,some work has been done on this problem based on the analysis and summary of the existing 3D face recognition algorithms,and two 3D face recognition methods are proposed.The main work and innovations of this paper are as follows:1.A 3D face recognition method based on multi-statistics local descriptor is proposed.First,the key points which are extracted in the semi-rigid region of the face based on the profiles are divided into two classes based on the sensitivity to expressions,and the neighborhood of these keypoints is determined at an appropriate value.Then,two order statistical information and first order statistical information in the neighborhood of keypoints are extracted using covariance matrix and histogram statistics,respectively,and then the feature-level fusion of the two is performed to obtain the multi-statistics local descriptor.Finally,those features of those keypoints which are less affected by facial expressions are given a higher weight and a multitask sparse representation classifier with weights is used to perform classification.The experimental results on the FRGC v2.0 and Bosphorus database show that the proposed algorithm is efficient and robust.2.A 3D face recognition method based on the mesh-VHLBP descriptor is proposed in this paper.First,the keypoints are extracted in the semi-rigid region of the face based on the profiles and the neighborhood of a keypoint is referred to the region constituted by the central facet and its surrounding concentric ordered rings.Then,the mesh-VLBP descriptor and mesh-HLBP descriptor are proposed respectively from vertical referring to the relation between the corresponding facets on two adjacent rings and horizontal referring to the relation between the facets with adjacent labels on one ring.Next,the feature-level fusion of the two descriptors is performed to obtain the mesh-VHLBP descriptor.Finally,the recognition experiments are conducted using Label Consist-KSVD2 learning algorithm on the FRGC v2.0 and Bosphorus database.The experimental results strongly demonstrate that the proposed method has a better recognition performance.