Research On Occluded Face Recognition Method Based On 3D Model

Automatic face recognition is one of the most promising areas of computer vision in many applications.The past two decades,since face recognition technology is non-invasive compared to other biometrics,it has become an active research branch in biometrics.Driven by a large number of potential applications,researchers continue to improve algorithms for more efficient.Although 2D face recognition has been thoroughly studied,the nature of human is three-dimensional objects.With the development of 3D measurement technology,3D data with more geometric information can be acquired more quickly and accurately,and more invariable features can be extracted for 3D recognition.Based on the matching algorithm of face point cloud and depth data,the three-dimensional face recognition of facial expression has been successfully applied,but few people have studied the problem of facial occlusion.In this paper,three-dimensional face recognition method is studied about the problem of facial occlusion.In order to maintain the details of face,this paper applies surface empirical mode decomposition on the grid to generate three-dimensional surfaces of different scales.On the surface,the key points extracted by the joint principal curvature are uniformly distributed on the face,and the reproducibility of the occurrence of the key points is improved.Using three kinds of descriptor to descript key points,and fused descriptor get more abundant information on local surface.A sparse representation matching algorithm with multiple dictionary cooperation is proposed.The specific research work on 3D face recognition in this paper is as follows:1)Extend the original Empirical Mode Decomposition(EMD)to the surface,and perform multi-scale decomposition on the face scan from 3D face database,reversing facial features.2)Joint principal curvatures on the surface to extract key points,establish a three-dimensional local coordinate system for each key point to extract rotation-invariant features.Reference on DAISY descriptors,constructed the structure containing more geometric features around key points.The shape index histogram,shape index gradient histogram,and surface gradient histogram,are used as descriptors.3)The occluded face extracts incomplete key points,which makes it difficult to match the similarity between points.This paper selects three different scales of face generated by decomposition,and uses the extracted key point descriptors to construct three sparse dictionaries respectively.The average reconstruction error of the descriptors is used as a measure of similarity,and a multi-dictionary collaborative sparse classification method is used to complete recognition.In this paper,we experiment in Bosphorus database using improved algorithm.we compared with other algorithms,our algorithm increased the accuracy of recognition,which has a certain positive impact on the future research of 3D face recognition.